Computing Performance Benchmarks among CPU, GPU, and FPGA

Transcription

Computing Performance Benchmarks among CPU, GPU, and FPGA
Computing Performance Benchmarks
among CPU, GPU, and FPGA
MathWorks
Authors:
Christopher Cullinan
Christopher Wyant
Timothy Frattesi
Advisor:
Xinming Huang
Abstract
In recent years, the world of high performance computing has been
developing rapidly. The goal of this project was to conduct computing performance
benchmarks on three major computing platforms, CPUs, GPUs, and FPGAs. A total of
66 benchmarks were evaluated. GPUs outperformed the other platforms in terms of
execution time. CPUs outperformed in overall execution combined with transfer
time. FPGAs outperformed for fixed algorithms using streaming. The team made
several recommendations for further research in this area.
i
Acknowledgements
The successful completion of this project would not have been feasible
without the help of several key individuals. First, we would like to express our
gratitude to our advisor, Xinming Huang, who met with us weekly throughout the
course of the project offering advice and wisdom. In addition, we would like to
thank our sponsor at MathWorks for their financial support of this project.
ii
Authorship
Christopher Cullinan
Christopher C. was responsible for the CPU Multicore portion of this project.
This included gathering and testing 34 benchmarks on the AMAX machine. In
addition, he wrote the CPU sections for the background, benchmark, and results
along with the future work and executive summary sections of this report.
Timothy Frattesi
Timothy was responsible for the GPU portion of this project. This included
gathering and testing 24 benchmarks on the GeForce GTX 460 and GeForce 9800
GTX+ NVIDIA graphics cards.
In addition, he wrote the GPU sections for the
background, benchmark, and results of this report along with the future work and
executive summary sections.
Christopher Wyant
Christopher W. was responsible for the FPGA portion of this project. This
included gathering and testing of 8 benchmarks using the ISim program of Xilinx on
a Virtex-5 board. Christopher W. wrote the FPGA sections of the background,
benchmark, and results section of the report. In addition, he was responsible for
writing the abstract, acknowledgements, introduction and conclusion of this report.
iii
Executive Summary
Ever since the beginning of modern day computing, engineers and developers have
been trying to squeeze every ounce of performance into their devices. How do we
test for performance though? What quantifiable measurements can be made to
justify the superiority of one device over another? These are the types of questions
this project aims to answer.
In this project, we investigated the performance
abilities of current generation multiprocessing hardware.
Through looking at
multicore CPUs, general purpose GPU computing and an FPGA, we compared device
capabilities to determine which platform future investments should be focused
towards and why.
To begin our endeavor, we used benchmarking to accentuate the strengths
and weaknesses of these three devices. Benchmarking is the technique of using
crafted programs in order to attach quantifiable performance metrics to targeted
computer subsystems. By using cross platform, as well as individual, benchmarks
developed across a plethora of computational necessities, we determined which
device would be best suited towards specific tasks. For this project, we tested our
benchmarks across two Intel Xeon 5650 CPUs, the Virtex-5 FPGA and NVIDIA’s
GeForce GTX460 and 9800 GTX+ GPUs.
Realizing the sophistication early on in this project, we decided to use
already written benchmarking suites to conduct our tests. A benchmarking suite is
nothing more than a compilation of individual benchmarks with specific intent. In
total, we used seven benchmarking suites. For the FPGA, we used cores designed by
Xilinx Core Generator and MATLAB Simulink HDL Coder, which contained
iv
benchmarks encompassing mathematical algorithms and encryption. For the CPUs,
we used three benchmarking suites; SPEC CPU2006, Rodinia, and John Burkardt.
Lastly, for the GPUs, we used the Parboil, Rodinia, and SHOC benchmarking suites.
The CPU and GPU suites tested mathematical algorithms, high performance
simulation, and common computational necessities such as compression and
sorting.
To further enhance the findings of this project, we discussed several future
recommendations at the end of this report. These recommendations include testing
a broader spectrum of benchmarks capable of running across all three platforms.
Another possibility is to look into newer technologies, such as an Accelerated
Processing Unit (APU). We believe that by including these recommendations, a
conclusion of greater impact can be reached.
v
Table of Contents
Abstract...................................................................................................................................................................... i
Acknowledgements .............................................................................................................................................. ii
Authorship .............................................................................................................................................................. iii
Executive Summary ............................................................................................................................................ iv
Table of Contents ................................................................................................................................................. vi
Table of Figures .................................................................................................................................................... ix
Table of Tables ....................................................................................................................................................... x
1.
Introduction .................................................................................................................................................. 1
2.
Background .................................................................................................................................................... 2
2.1.
2.1.1.
A Brief History ............................................................................................................................. 2
2.1.2.
CPU Design .................................................................................................................................... 4
2.1.3.
CPU Multicore .............................................................................................................................. 6
2.2.
Graphics Processing Unit.................................................................................................................. 9
2.2.1.
GPU History ................................................................................................................................... 9
2.2.2.
GPU Architecture and Parallelism ..................................................................................... 12
2.3.
3.
CPU ............................................................................................................................................................ 2
FPGA Background.............................................................................................................................. 17
2.3.1.
History of FPGAs ...................................................................................................................... 17
2.3.2.
Early Programmable Devices .............................................................................................. 18
2.3.3.
FPGA Architecture ................................................................................................................... 19
Benchmarks ................................................................................................................................................ 24
3.1.
FPGA Benchmarks ............................................................................................................................. 26
3.2.
SPEC CPU2006 .................................................................................................................................... 28
3.2.1.
CPUINT2006 .............................................................................................................................. 28
3.2.2.
CFP2006 ...................................................................................................................................... 34
3.3.
Rodinia Suite ....................................................................................................................................... 41
3.4.
John Burkardt Benchmarks ........................................................................................................... 44
3.5.
SHOC Suite............................................................................................................................................ 45
3.6.
Parboil Suite ........................................................................................................................................ 48
vi
Breadth ......................................................................................................................................................... 48
4.
Results .......................................................................................................................................................... 52
4.1.
Devices ................................................................................................................................................... 52
4.1.1.
GPU ................................................................................................................................................ 52
4.1.2.
CPU ................................................................................................................................................ 53
4.1.3.
FPGA ............................................................................................................................................. 53
4.2.
ALL Devices.......................................................................................................................................... 54
4.3.
CPU & GPU ............................................................................................................................................ 55
4.4.
Individual Results.............................................................................................................................. 57
4.4.1.
GPU Specific ............................................................................................................................... 57
4.4.2.
CPU Results ................................................................................................................................ 65
4.4.3.
FPGA Results.............................................................................................................................. 72
5.
Future Work ............................................................................................................................................... 74
6.
Conclusion ................................................................................................................................................... 75
7.
Appendices.................................................................................................................................................. 76
7.1.
Parboil Results 9800 GTX+ ............................................................................................................ 76
7.2.
Parboil Results GTX 460 ................................................................................................................. 78
7.3.
Rodinia Results 9800 GTX+ ........................................................................................................... 80
7.4.
Rodinia Results GTX 460 ................................................................................................................ 81
7.5.
SHOC Max Flops GTX 460 ............................................................................................................... 84
7.6.
SHOC Bus Download Speed GTX 460......................................................................................... 85
7.7.
SHOC Device Memory GTX 460 .................................................................................................... 86
7.8.
SHOC SPMV GTX 460........................................................................................................................ 87
7.9.
SHOC MD GTX 460 ............................................................................................................................ 88
7.10.
SHOC Reduction GTX 460 .......................................................................................................... 89
7.11.
SHOC S3D GTX 460 ....................................................................................................................... 90
7.12.
SHOC Scan GTX 460 ...................................................................................................................... 91
7.13.
SHOC SGEMM GTX 460 ............................................................................................................... 92
7.14.
SHOC Sort GTX 460 ....................................................................................................................... 93
7.15.
SHOC Stencil 2D GTX 460........................................................................................................... 94
7.16.
SHOC Triad GTX 460 .................................................................................................................... 95
vii
7.17.
SHOC Max Flops 9800 GTX+ ..................................................................................................... 96
7.18.
SHOC Bus Download Speed 9800 GTX+ ............................................................................... 97
7.19.
SHOC SPMV 9800 GTX+ .............................................................................................................. 98
7.20.
SHOC MD 9800 GTX+ ................................................................................................................... 99
7.21.
SHOC Reduction 9800 GTX+ .................................................................................................. 100
7.22.
SHOC S3D 9800 GTX+ ............................................................................................................... 101
7.23.
SHOC SGEMM 9800 GTX+ ....................................................................................................... 102
7.24.
SHOC Sort 9800 GTX+ .............................................................................................................. 103
7.25.
SHOC Stencil 2D 9800 GTX+ .................................................................................................. 104
7.26.
SHOC Triad 9800 GTX+ ............................................................................................................ 105
7.27.
SPEC CPU2006 Integer Results (No Auto-Parallel) ...................................................... 106
7.28.
SPEC CPU2006 Integer Results (Auto-Parallel Enabled) ........................................... 106
7.29.
Speedup of SPEC CPU2006 Integer Results ..................................................................... 107
7.30.
SPEC CPU2006 Floating Point Results (No Auto-Parallel)......................................... 107
7.31.
SPEC CPU2006 Floating Point Results (Auto-Parallel Enabled) .............................. 108
7.32.
Speedup of SPEC CPU2006 Floating Point Results ....................................................... 108
7.33.
Rodinia/Burkardt Benchmarks Average Execution Times ....................................... 109
7.34.
Rodinia/Burkardt Benchmarks Speedup between Thread Count .......................... 109
7.35.
FPGA Results ................................................................................................................................ 110
Bibliography....................................................................................................................................................... 111
viii
Table of Figures
Figure 1 - AMD K10 Architecture ..................................................................................................... 6
Figure 2 - Intel I7 950 Quad Core Processor Design ................................................................. 7
Figure 3 - S386C911 (1991) ............................................................................................................ 10
Figure 4 - NVIDIA GeForce 256 (1998) ....................................................................................... 10
Figure 5 - GeForce 6600 (2004) ..................................................................................................... 11
Figure 6 - GeForce GTX 560 (2011) .............................................................................................. 11
Figure 7 - CUDA Software Architecture ...................................................................................... 12
Figure 8 - Improvements in CUDA GPU Architecture ............................................................ 14
Figure 9 - Thread Hierarchy ............................................................................................................ 16
Figure 10 - Stream Multiprocessor ............................................................................................... 16
Figure 11 - PLA Architecture ........................................................................................................... 18
Figure 12 - CPLD Architecture ........................................................................................................ 19
Figure 13 - Three Input LUT ............................................................................................................ 20
Figure 14 - Switch Block.................................................................................................................... 21
Figure 15 - FPGA Routing ................................................................................................................. 22
Figure 16 - Specialized Slices .......................................................................................................... 23
Figure 17 – Virtex-7 Architecture.................................................................................................. 23
ix
Table of Tables
Table 1 - FFT Results .......................................................................................................................... 54
Table 2 - CPU & GPU Results............................................................................................................ 56
Table 3 - GeForce Specification ...................................................................................................... 58
Table 4 - Data Size vs. Data Rate .................................................................................................... 59
Table 5 – Global vs. Local Memory Speeds ................................................................................ 60
Table 6 – Parboil Suite Timings ...................................................................................................... 63
Table 7 - Rodinia Particle Filter Results .................................................................................... 64
Table 8 - SPEC CPUINT2006 w/ GCC, G++, GFortran compiler w/o auto parallel ..... 65
Table 9 - SPEC CPUINT2006 run with Intel compiler in auto parallel ............................ 66
Table 10 - Speedup percentages from the GNU run to the Intel run ............................... 66
Table 11 - SPEC CFP2006 w/ GCC, G++, GFortran compiler w/o auto parallel ........... 68
Table 12 - SPEC CFP2006 run with Intel compiler in auto parallel ................................. 68
Table 13 - Speedup percentages from the GNU run to the Intel run ............................... 69
Table 14 - Rodinia/Burkardt average execution time on 2, 4, 8, 12, 24 threads ........ 70
Table 15 - Speedup between thread counts .............................................................................. 71
Table 16 - FPGA Results .................................................................................................................... 72
x
1.
Introduction
The world of high performance computing is a rapidly evolving field of study. Many
options are open to businesses when designing a product.
GPUs can provide
astonishing performance using the hundreds of cores available. On the other hand,
FPGAs can provide computational acceleration to many signal and data processing
applications. The question arises as to what level each platform performs at for
different benchmark algorithms.
To determine computing levels of each device we implemented existing
benchmark suites for each device. We tested several applications to see which
computing method was fastest for the various applications. While the benchmarks
did not completely lineup between the different processors the information
gathered laid a good foundation between different devices.
In order to explain the information in a clear manner, we broke up the
information into several sections. Presented first is the background of the devices,
both general and specific. The second section outlines all of the benchmarks we
tested so that the reader can understand the limitations of the project. The third
section discusses the results we gathered as well as a discussion of what they mean.
Lastly, we discuss some recommendations we would make for similar projects in
the future and some closing remarks.
1
2.
Background
The background gives an overview of the different device we used in our project.
Within each section are the history of the device and a general overview of how they
work. This information gives a brief but necessary background into the platforms
used for this project.
2.1.
CPU
2.1.1. A Brief History
The history of the Central Processing Unit (CPU) is in all respects a relatively
short one, yet it has revolutionized almost every aspect of our lives. In the early
1970, if I were to ask someone what a CPU was, they would have most likely
responded “A what!” Yet just over 40 years later, CPUs have become an integral part
of our lives. From desktop computers to cell phones, most of us do not go more than
a few hours without somehow interacting with a CPU. Despite its indisputable
popularity, most do not know how this hype all started.
In 1971, the 4-bit Intel 4004 was the first in the legacy of the CPU. It was the
first commercially available CPU on chip, made possible by the all-new silicon gate
technology. The max CPU clock rate of this revolutionary hardware was 740 kHz, an
astonishing speed at the time. This little guy could execute 92,000 instructions per
second with a single instruction cycle of 10.8 microseconds and a transistor count of
2,300. At the time, this device was truly a feat in computing technology, which
paved the road for much more innovation to come. [1]
2
Intel dominated CPU infancy, coming out with several subsequent CPU
designs including the 8086 (1978), 8088 (1979), 80186 (1980). Then, in 1993, one
of the most popular names in the history of the CPU surfaced, the Intel Pentium
processor. This legendary device operated at a whopping 60 MHz and 100 Millions
of Instructions per Second (MIPS). The trend of innovation from Intel continued for
several years until another major competitor in today’s market made their first
competitive appearance with the AMD AM5x86 in 1995.
A fierce competition
between Intel and AMD has continued since. [2]
The next milestone in the CPU history was the commercial release of the first
1GHz processor. This achievement was reached by the AMD Athlon in 1999 and
then by the Intel Pentium III just two days later after. For this reason, “Athlon” was
fitting name for AMD’s milestone processor because it is the Greek word for
“Champion/trophy of the games”. The AMD Athlon is an x86-compatible processor
containing 22 million transistors in a slim size of 184 mm2. [3]
Nowadays, it is a common occurrence to see CPUs clocked well above 1GHz
in devices as small as our cell phones. In just over 40 years, we have gone from 740
kHz to the GHz level (over a 1300 % increase) and increased the count of on chip
transistors from 2,300 to more than a billion (over a 434,000 % increase). We are
now producing CPUs with multiple cores on the same chip, which are capable of
support an increasingly important feature known as parallel computing, which we
will talk about in more detail later in this report. [2]
3
2.1.2. CPU Design
In a nutshell, CPU design is the design engineering process of creating a
central processing unit to be used in a computing system. Many factors go into
designing a CPU, especially with the level of sophistication in modern day CPUs.
There are six primary focuses that designers must account for when creating a CPU,
and they are; data paths, control unit, memory components, clock circuitry, pad
transceiver circuitry, and logic gate cell library. [4]
A data path by definition is, “A collection of functional units, such as
arithmetic logic units or multipliers, that perform data processing operations”. [5]
Intuitively from its name, data paths provide routes for data to traverse between the
components of a CPU. These routes are typically known as “Buses”. The majority of
CPUs include both a data path and a control unit, where the control unit specializes
in regulating data path and main memory interaction. [5]
Most modern day CPUs have several types of memory modules on chip. Two
of the most popular are register memory and cache, both of which are normally high
speed SRAM. Registers are the memory cells built directly into the CPU since they
contain specific data vital to CPU operations. Cache is the next portion of memory in
a CPU and is usually, in more complex processors, divided into L1 (level one) and L2
(level two) cache. Both L1 and L2 cache are there to store data that is most often
used by the CPU and is typically SRAM as well. [6]
The CPU clock is the sinusoidal frequency reference signal typically created
by a crystal oscillator. This sinusoidal waveform is first translated into a square
waveform of the same frequency by internal circuitry and then used to synchronize
4
the internal components of the CPU. The clock signal traverses to the various CPU
components via a clock distribution network. [7]
Lastly, the logic gate cell library is the collection of logic gates used to
implement computational logic in the CPU. The logic library collection consists of
low-level logic functions including AND, OR, INVERT gates as well as flip-flops,
latches, and buffers. A vital feature of these libraries is that they are fixed height and
variable width, meaning they can be placed in organized rows. This makes the
process of automated digital layout of these components possible and efficient. [8]
To give you an idea of a simple CPU design, the following figure shows the
AMD K10 Architecture of 2007. This processor is slightly outdated but is good for
our purposes to show the anatomy of a modern day CPU.
5
Figure 1 - AMD K10 Architecture
2.1.3. CPU Multicore
The term “multi-core” refers to a multiple core processor that is simply an
integrated circuit where two or more processors have been attached for increased
performance via parallel processing. [9] Parallel processing is a type of computation
6
where many calculations are performed simultaneously.
This method of
computation is based on the principle that large problems can be solved faster by
breaking them down into smaller pieces and then solving those pieces concurrently.
Because of this basic principle, parallel computing has become the dominant
standard in computer architecture in the most popular form of multicore
processing. As an example of a multicore processor design, the following figure
shows the internals to an Intel I7 950 Quad Core Processor.
Figure 2 - Intel I7 950 Quad Core Processor Design
In the real world, parallel processing is not limited to integrated circuits.
Virtually everything in our natural universe uses the principles of parallel
processing. A few examples are galaxy formation, planetary movements, weather
and ocean patterns, automobile assembly lines, rush hour traffic, and even ordering
a hamburger at a fast food restaurant. Within all of these phenomena, numerous
7
complex and interrelated events occur simultaneously to achieve a common goal.
[10]
In the past, parallel computing was attributed mainly to high end computing
and was used to solve complex mathematical problem in various areas of study. A
few of these areas included Atmospheric studies, Physics, Mechanical Engineering,
Electrical Engineering, and Seismology. Parallel computing is still used in those
areas today, but the rise of commercial applications has been a major contributor to
both the need for faster computing and the dispersion of parallel computing into
common electronic devices such as phones, desktops, laptops, etc. Database mining,
web search engines, medical imaging, and advanced graphics are just a few of the
applications that utilize parallel computing. [10]
At the beginning of this section, we briefly discussed the incentives to use
parallel computing; now we will delve deeper to justify the use of parallel
computing. Parallel computing saves time and money by both shortening the time
to the outcome and because parallel components are cheap.
Second, larger
problems can be solved through the use of parallel computing that are not possible
by using a single computing resources. Finally, there are many limitations to serial
computing. These limitations include transmission speeds, limits to miniaturization,
and economic limitations. To get away from these confines, modern computer
architectures are heavily relying on multiple execution units, pipelined instructions,
and multi-core at the hardware level to increase performance. [10]
8
2.2.
Graphics Processing Unit
As stated by Prof. Jack Dongarra, "GPUs have evolved to the point where
many real-world applications are easily implemented on them and run significantly
faster than on multi-core systems. Future computing architectures will be hybrid
systems with parallel-core GPUs working in tandem with multi-core CPUs." [11]
From this comment one can see one of the many thoughts on where the future or
high performance computing is headed.
2.2.1. GPU History
Much like computers in general, GPU's have progressed rapidly over the last
30 years since their introduction to the market. As GPU's have progressed over the
years their core functions have remained the acceleration and processing of images.
The introduction of graphics units came early in the 1980's where both Intel
and IBM brought specialized products to the market. Other companies such as
Commodore and Texas Instruments also added simple graphics capabilities either
on chip or using an external card. These cards had simplistic functionality and were
relatively expensive. Functions such as filling an area, shape drawing, and
modification of simple images were all that these early processors could support.
9
Figure 3 - S386C911 (1991)
Figure 4 - NVIDIA GeForce 256 (1998)
The 1990's were the real beginning as far as the takeoff of GPUs. From the
beginning in 1991, S3 rolled out their 86C911 card which was one of the first
standards for the GPU industry. Two dimensional graphic processing had made its
way into almost every system by the mid 90's and the race was on to move towards
3D processing. Two notable chip sets in the race for dedicated 3D graphics include
the 3dfx Voodoo and Rendition's Verite. Until the late 90's all 3D rendering was
done with the assistance of CPUs, also known as hardware assisted 3D graphics
which we still see in lower end laptops today. [12]
To assist in the commonality of graphics processing several “languages” were
brought about in the late 90's including both OpenGL and Direct. Throughout the
90's OpenGL prospered as the software’s capability was usually ahead of Direct and
it was capable of being used across cards and platforms. Towards the end of the
90's these two API's introduced support for transform and lighting (T&L) which
provided a huge jump in GPU processing. T&L allowed for easier mapping of 3D
images to a 2D plane while incorporating the lighting all into one. By this time there
were only a few competing companies; NVIDIA, ATI, 3dfx, and S3. The end of the
10
90's saw the NVIDIA GeForce 256, the first readily available commercial card,
bringing 3D graphics, NVIDIA, and Direct to their own level. [12]
Figure 5 - GeForce 6600 (2004)
Figure 6 - GeForce GTX 560 (2011)
Through 2010 and to today we continue to see significant advancements in
the 3D rendering abilities of the GPU. On the front of programming the most notable
improvements include programmable shading and floating point abilities. ATI and
NVIDIA hold the majority of today’s market share in graphics processing and thus
have been major forces in shaping how these units improve.
One of the most significant advancements of the past decade is general
purpose computing for GPU’s. Due to the highly parallel structure of modern
graphics cards it is possible to use them to perform research and analysis, often
times competing or surpassing modern CPUs. While this can be done with almost
any modern card, NVIDIA’s introduction of the Compute Unified Device Architecture
(CUDA) from NVIDIA this idea has become standardized. OpenCL is also a common
language for performing GPU computation, but it does not support as many
programming languages or have the same amount of industry support. CUDA
11
architecture is the main advancement that is allowing our society to take high
performance computing from CPUs and FPGAs and move it to a quicker paralleled
set of computations on thousands of threads instead of tens of threads. [13]
2.2.2. GPU Architecture and Parallelism
Here we will look at the both the CUDA architecture and the hardware
architecture that corresponds to it. NVIDIA has created this specialized architecture
to achieve the massively parallel systems that we have today.
Figure 7 - CUDA Software Architecture
Looking backwards from computer to device we start with the Parallel
Thread Execution (PTX) instruction set. This specially optimized instruction set
allows for special optimization specifically designed for parallel processing on
NVIDIA cards. The PTX instruction set allows NVIDIA to set a standard across
multiple generations of GPUs as well as provide a common set of instructions for
12
both optimization and developer programming. This instruction set rests within the
CUDA drivers which are provided dating back to NVIDIA’s GeForce 8800 series. [14]
The next level up, CUDA Support in OS is provided through the CUDA Toolkit
which is currently on version 4.1.
This toolkit provides all of the necessary
components to write and run CUDA code in an IDE such as Visual Studio or simply in
a text editor. Currently supported are both a LLVM based and the standard nvcc
compiler. The move to the LLVM based compiler has shown 10% increases in
speed, but is geared more towards the new Fermi architecture and beyond. With
the release of the new version of the CUDA toolkit, NVIDIA also provided a much
larger base of functions for image processing.
The base level is the actual CUDA Parallel Computing engines that are the
basis of all modern NVIDIA cards. The specialized hardware architecture in these
cards is what allows for such successful parallelism in general purpose computing.
On the current generation of CUDA capable hardware, Fermi, one can setup and
process 65,535 simultaneous threads in grids of up to 1024x1024x64. These grid
sizes have doubled and allowed for a third dimension since CUDAs initial release in
2006.
The number of cores capable of providing floating point and integer
functionality has also increased six fold. In the following figure the hardware
advancements can be seen in overview as the architecture has changed. Further
improvements can all be seen in the tables attached in Appendix X. [15] [16]
13
Figure 8 - Improvements in CUDA GPU Architecture
Parallelism is generated through the use of threads on the GPU. CUDA has
the ability to run a particular kernel across multiple threads and multiple cores.
Picture the newest card in NVIDIA’s lineup has 1024 cores per card. If each core can
provide up to 1024 threads per 65535 blocks, there is a possibility to run upwards
of 60 billion instances of a single kernel per card in a system. Spanning this number
across multi-card systems, or even multiple systems connected together, it is easy to
see how parallelism prevails and allows for faster processing.
This type of
parallelism is known as single instruction multiple data (SIMD). CUDA devices are
also capable of running thousands of small programs simultaneously as well.
In order to achieve this number of parallel threads there is a complex setup
of memory. Each and every thread an individual piece of memory that is unshared.
This contains data such as program counters and individual registers. From there
we move up to thread blocks which share memory amongst themselves and then up
to sixty five thousand blocks sharing memory per core. These sets of blocks are
14
called grids and can share a set of application memory for use by smaller threads
with global memory. All of the threads running a particular instance of a kernel are
kept in synchronization through the use of special code functions that wait for all
threads to be completed before reading or writing large changes from global
memory. The memory hierarchy described here is important because each smaller
level allows for quicker access to data; thus like the L cache of a CPU, there is less
need to constantly read and write to slower global memory. [16] [17]
To handle all of these threads processing at the same time there is a unique
feature called a warp handler. A warp is a group of 32 threads; the smallest data
size for a SIMD setup. When programming in CUDA, users work with blocks so it is
up to the warp handler to determine how to divide the instructions. The Fermi
architecture has a dual warp scheduler, allowing it to process and divide up two sets
of instructions at a time. With 32 bit mathematics it is also possible to dispatch two
of a single type of instruction or a mix at one time. Since this setup works in sets of
32 in order to achieve the peak performance on CUDA capable GPUs is to run
kernels in sets divisible by 32. [17]
Along with the efficient thread hierarchy, the Fermi architecture relies on
NVIDIA’s third generation of stream multiprocessing (SM) for its hardware
architecture. In this revision of SM, there are 32 CUDA cores per multiprocessor,
giving each card 16 to 32 SMs; with each core having it’s out floating point and
arithmetic logic units. Figure 9 below is a simplified example of third generation
SM. Pictured is a single multiprocessor, with the hardware available in each SM. In
the figure the 32 individual processing cores can be seen along with the 16
15
load/store units and the 4 special functions units. These units are accessed once the
warp schedulers divide tasks amongst the cores. Following these processing units
are the standard graphical processing hardware units such as those that perform
tessellations and texturing. It is also noted that in this version each individual core
has its own integer and floating point units. [16] [17]
Figure 9 - Thread Hierarchy
Figure 10 - Stream Multiprocessor
Added support for the new IEEE floating point standard has given these new
cards the ability to use a fused add and multiply in one step. Data precision has also
been improved so that the integer units provide 64 bit support while the floating
point units finally provide full 32 bit support. The sixteen load and store and four
special function units allow for up to sixteen thread address to be calculated at a
16
time and four instances of functions such as sine or square roots. With most of these
instructions executing with one instruction per clock cycle having so many SMs on a
single card allows for distribution of complex equations and faster execution. [16]
2.3.
FPGA Background
2.3.1. History of FPGAs
Ross Freeman, co-founder of the company Xilinx, invented Field
Programmable Gate Arrays (FPGA) in 1894 while working for the company Zilog.
After inventing the FPGA Freeman left Zilog with his Patent (patent 4,870,302) to
found Xilinx. While Xilinx is a multi-billion dollar company today, Freeman did not
live to see this become a reality passing away in 1989. He was honored in 2009 by
being inducted into the National Inventor’s Hall of Fame for his work on FPGAs. [18]
The first FPGAs released to the market had only several thousand gates and
had several disadvantages to their counterparts, ASICs.
They were slower,
consumed more power, and had limited functionality. The industry of FPGAs grew
slowly through the 1990s. In 1992 the U.S. Naval Surface Warfare department
completed a project on FPGAs that implemented 600,000 logic gates. During this
time, the main applications for FPGAs were networking and telecommunications.
By the late 90s, the number of gates on a single FPGA reached the millions
and many of the disadvantages compared to ASICs were diminishing. FPGAs began
entering many other industries because of the low time from development to
market introduction. Much money could be generated by being the first to the
market. [19]
17
Today, FPGAs can cater to many different applications. Different series and
families are application specific and have additional logic to support faster
processes. FPGAs have a high capacity for parallelization and pipelining processes.
Often, they are used as peripherals to CPUs to carry out specific processes that a
CPU has trouble handling.
2.3.2. Early Programmable Devices
The idea behind FPGAs originated from two devices, Programmable Logic
Arrays (PLA) and Complex Programmable Logic Devices (CPLDs).
PLAs were
introduced during the early 1970s as one- time programmable chips to implement
logic functions. The AND gates and OR gates were connected with a communication
matric that could be programmed by burning fuses to implement a truth table. The
limiting factors were the number of inputs, AND gates, and OR gates. [20]
Figure 11 - PLA Architecture
CPLDs built upon the idea of PDAs with an interconnection matrix connecting
all of the inputs and outputs. The connection matrix was formed of on-chip Flash
18
memory to configure macrocells. These macrocells are very similar in structure to a
PLA. CPLDs are very similar to FPGAs as the main difference is in the underlying
architecture. [20]
Figure 12 - CPLD Architecture
2.3.3. FPGA Architecture
Field Programmable Gate Array is a semiconductor device comprised of
many logic blocks with configurable interconnections between these. The logic
blocks are capable of acting as simple logic gates, such as AND and XOR. In addition
to the logic gates, there are routing channels that run between each logic block.
These channels are programmable and enable different logic blocks to talk to each
other.
In recent years, more specific circuits are implemented on FPGAs for
19
application specific purposes. These can include multipliers and DSP circuits, which
speed up processing for those applications.
The main component of FPGAs is the logic block. Millions of these are
replicated in a network throughout the chip. They are implemented in a Lookup
Table (LUT) usually consisting of four input pins. The LUTs have small piece of
memory attached that is programmed for output logic depending on the input.
Essentially, a truth table defined for that piece of logic.
Some designers are
increasing the number of input pins to six to increase speed.
Figure 13 - Three Input LUT
Each LUT has only one output. This output can then be stored in a flip flop to
preserve values over a clock cycles, or it can run to other LUTs to further implement
logic. The Virtex-5, which the benchmarks in the report are based on, uses six input
LUTs.
20
The routing channels, which run between logic blocks, are used to connect
various LUTs together. The routing channels are controlled by switch blocks that
control connections between crossing wires. An example of this is seen below in
Figure 14.
Figure 14 - Switch Block
These connections allow for an immense amount of configurable logic
allowing an FPGA to carry out its functionality. Figure 15 below shows the layout of
blocks throughout a network. [21]
21
Figure 15 - FPGA Routing
In newer FPGA series other blocks are implemented for application specific
functionality. These specialized blocks, or slices, run their specific functionality
faster than can be implemented using LUTs and routing wires. These slices also
drastically cut down the number of LUTs used. The two main blocks are the
multipliers and DSP slices. To implement the multiplication of two 32 bit numbers
would require more than 2000 operations for a single multiply. Different FGPA
series have different types and quantities of specialized blocks based on designed
applications. [22]
22
Figure 16 - Specialized Slices
The latest FPGA is the Virtex-7 from Xilinx.
This model has increased
computing power and efficiency. The architecture for this model can be seen below
in Figure 17.
Figure 17 – Virtex-7 Architecture
23
3.
Benchmarks
Benchmarking has been around for years and is widely used as the standard to
which we base our computing processing power today. Programs available from
many different packages are used to measure the metrics of new processors daily.
In the beginning, performance was measured by various system specifications, such
as clock rate. Many designers and consumer realized that while this may give some
indication of the processing power of a machine, it did not incorporate the entire
scope of the situation.
Many benchmarks are used today to apply stress to
processors through different processes for different applications.
Benchmarks
come in four different types, each having its own strength and weaknesses: real
applications, small benchmarks, benchmark suites, and synthetic benchmarks.
Real applications are benchmarks bases pre-existing programs. They are
comprised of a typical user’s workload during the day. The advantages to running
real applications are that they directly translate over to improved performance
times on programs and very accurately mirror everyday workloads. However, these
benchmarks are usually very big and require more time to run and transfer to other
machines. In addition, it can be very hard to pinpoint a processing bottleneck to
discover what instruction types need to be improved for the greatest performance
increase.
The next type of benchmark is the small benchmark. It consists of a very
small code segment that exists in many other applications.
For example, the
following C code could be a small benchmark.
24
for (j = 0; j<8; j++)
S = S + Aj  Bi-j;
This code runs very fast and does not take very long to compile or transfer
between systems. In addition, it can give developers a very good idea about what
portion of their process is bottlenecking the systems in order to make
improvements and can be easy to simulate during design to test functionality.
However, many designers can take advantage of the limited instructions used to
design a system specialized for that particular loop. This abuses the benchmark and
does not report accurate data to both designers and customers.
Benchmark suites are compilations of different benchmarks from different
industries that together represent a variety of computing loads on a machine.
Standard Performance Evaluation Corporation (SPEC) provides one of the main
suites available to developers. SPEC began in 1989 with SPEC89 CPU intensive
benchmark suite. Many companies came together and agreed on a set of programs
that represented a user’s typical workload. Suites are very useful in covering a
diverse set of parameters and characteristics; however, they are still susceptible to
abuse. In addition, they require periodic updates to change the applications as
typical workloads change.
Lastly, some programmers advocate the use of synthetic benchmarks. These
programs attempt to mirror the characteristics of other applications while using
much less space and processing time. In reality, they do not perform any functional
task on the processor, but do give an accurate representation of processing power.
25
It is always important to run many iterations of a benchmark. Timings will
change and an average of all iterations should be used to provide a more accurate
picture for comparisons. Often manufacturers can find one or more benchmarks
that their platform particularly excels at and abuse this standard for advertisement.
The best benchmarks to look at for different computers are those posted by nonprofit organizations which are unbiased. [23]
3.1.
FPGA Benchmarks
Benchmarking is not traditionally done for FPGAs. Rather, the datasheets for
different FPGAs have details about maximum clock cycle that they can operate at.
The user then looks at the program that they want to run and they can see number
of clock cycles it takes to produce a result. The throughput does not change
between iterations on the same machine. The following benchmarks that were run
on the Virtex-5 series FPGA were compiled using either the Xilinx Core Generator or
MATLAB Simulink HDL Coder.
The Xilinx CORE Generator is where most of the applications are from. It is
built into the Xilinx program. When opening a core you have several options that
can be selected based on your needs such as extra pin I/O, processing type, and size.
The Simulink HDL coder is a new addition to the program. Rather than designing
our own code, the current demos available in the Simulink program were used to
generate HDL code, which is then tested in the Xilinx program. The HDL Coder is
also capable of generating testbenches for the application.
26
Fast Fourier Transform (FFT)
A discrete Fourier transform that operates with reduced computational power. The
number of computations needed is reduced from 2N^2 to 2N Log2 N where N is the
number of points necessary for the computation. The FFT used in this study is a
1024 point FFT.
Finite Impulse Response Filter (FIR)
FIR Filters are designed to be simple to implement for Digital Signal
Processing (DSP). FIR filters are more commonly used then IIR filters because of the
advantages offered such as fractional arithmetic and fewer practical problems. The
FIR filter used for performance testing in this study is a low pass filter.
Advanced Encryption Standard (AES)
The Advanced Encryption Standard (AES) specifies a FIPS-approved
cryptographic algorithm that can be used to protect electronic data. The AES
algorithm is a symmetric block cipher that is capable of encrypting (encipher) and
decrypting (decipher) information. Encryption converts data to an unintelligible
form called cipher text; decrypting the cipher text converts the data back into its
original form, called plaintext. The AES algorithm is capable of using cryptographic
keys of 128, 192, and 256 bits to encrypt and decrypt data in blocks of 128 bits.
(http://csrc.nist.gov/publications/fips/fips197/fips-197.pdf)
Double Precision Floating Point Multiplication
27
This block takes in two double floating point numbers for multiplication.
This is similar to a small benchmark for CPUs. A single FPGA is able to implement
multiple instances of this application allowing for large amounts of parallel
processing.
3.2.
SPEC CPU2006
Standard Performance Evaluation Corporation (SPEC) CPU 2006 is a
benchmark suite that contains over twenty-five different benchmarks.
These
industry-standardized benchmarks were created to stress a system’s processor,
memory architecture, and compilers. This particular suite is divided into two sub
categories, CINT2006 and CFP2006. The CINT2006 portion consists of twelve total
benchmarks that measure integer operation performance, while the CFP2006
portion consists of seventeen benchmarks that measure floating point operation
performance. A brief description of each SPEC CPU2006 benchmark can be seen
below. [24]
3.2.1. CPUINT2006
400.perlbench
“400.perlbench is a cut-down version of Perl v5.8.7, the popular scripting
language. SPEC's version of Perl has had most of OS-specific features removed. In
addition to the core Perl interpreter, several third-party modules are used:



SpamAssassin v2.61
Digest-MD5 v2.33
HTML-Parser v3.35
28




MHonArc v2.6.8
IO-stringy v1.205
MailTools v1.60
TimeDate v1.16” [25]
401.bzip2
“401.bzip2 is based on Julian Seward's bzip2 version 1.0.3. The only
difference between bzip2 1.0.3 and 401.bzip2 is that SPEC's version of bzip2
performs no file I/O other than reading the input. All compression and
decompression happens entirely in memory. This is to help isolate the work done to
only the CPU and memory subsystem.” [25]
403.gcc
“403.gcc is based on gcc Version 3.2. It generates code for an AMD Opteron
processor. The benchmark runs as a compiler with many of its optimization flags
enabled.
403.gcc has had its inlining heuristics altered slightly, so as to inline more
code than would be typical on a UNIX system in 2002. It is expected that this effect
will be more typical of compiler usage in 2006. This was done so that 403.gcc would
spend more time analyzing its source code inputs, and use more memory. Without
this effect, 403.gcc would have done less analysis, and needed more input workloads
to achieve the run times required for CPU2006.” [25]
429.mcf
29
“429.mcf is a benchmark that is derived from MCF, a program used for singledepot vehicle scheduling in public mass transportation. The program is written in C.
The benchmark version uses almost exclusively integer arithmetic.
The program is designed for the solution of single-depot vehicle scheduling
sub-problems occurring in the planning process of public transportation companies.
It considers one single depot and a homogeneous vehicle fleet. Based on a line plan
and service frequencies, so-called timetabled trips with fixed departure/arrival
locations and times are derived. Each of these timetabled trips has to be serviced by
exactly one vehicle. The links between these trips are so-called dead-head trips. In
addition, there are pull-out and pull-in trips for leaving and entering the depot.” [25]
445.gobmk
“The program plays Go and executes a set of commands to analyze Go
positions.” [25]
456.hmmer
“Profile Hidden Markov Models (profile HMMs) are statistical models of
multiple sequence alignments, which are used in computational biology to search
for patterns in DNA sequences.
The technique is used to do sensitive database searching, using statistical
descriptions of a sequence family's consensus. It is used for protein sequence
analysis.” [25]
30
458.sjeng
“458.sjeng is based on Sjeng 11.2, which is a program that plays chess and
several chess variants, such as drop-chess (similar to Shogi), and 'losing' chess.
It attempts to find the best move via a combination of alpha-beta or priority
proof number tree searches, advanced move ordering, positional evaluation and
heuristic forward pruning. Practically, it will explore the tree of variations resulting
from a given position to a given base depth, extending interesting variations but
discarding doubtful or irrelevant ones. From this tree the optimal line of play for
both players ("principle variation") is determined, as well as a score reflecting the
balance of power between the two.
The SPEC version is an enhanced version of the free Sjeng 11.2 program,
modified to be more portable and more accurately reflect the workload of current
professional programs.” [25]
462.libquantum
“Libquantum is a library for the simulation of a quantum computer. Quantum
computers are based on the principles of quantum mechanics and can solve certain
computationally hard tasks in polynomial time. In 1994, Peter Shor discovered a
polynomial-time algorithm for the factorization of numbers, a problem of particular
interest for cryptanalysis, as the widely used RSA cryptosystem depends on prime
factorization being a problem only to be solvable in exponential time. An
implementation of Shor's factorization algorithm is included in libquantum.
31
Libquantum provides a structure for representing a quantum register and
some elementary gates. Measurements can be used to extract information from the
system. Additionally, libquantum offers the simulation of decoherence, the most
important obstacle in building practical quantum computers. It is thus not only
possible to simulate any quantum algorithm, but also to develop quantum error
correction algorithms. As libquantum allows adding new gates, it can easily be
extended to fit the ongoing research, e.g. it has been deployed to analyze quantum
cryptography.” [25]
464.h264ref
“464.h264ref is a reference implementation of H.264/AVC (Advanced Video
Coding), the latest state-of-the-art video compression standard. The standard is
developed by the VCEG (Video Coding Experts Group) of the ITU (International
Telecommunications Union, http://www.itu.int) and the MPEG (Moving Pictures
Experts
Group, http://mpeg.chiariglione.org)
of
the
ISO/IEC
(International
Standardization Organization, http://www.iso.ch). This standard replaces the
currently widely used MPEG-2 standard, and is being applied for applications such
as the next-generation DVDs (Blu-ray and HD DVD) and video broadcasting.” [25]
471.omnetpp
“The benchmark performs discrete event simulation of a large Ethernet
network. The simulation is based on the OMNeT++ discrete event simulation system
(www.omnetpp.org), a generic and open simulation framework. OMNeT++'s
32
primary application area is the simulation of communication networks, but its
generic and flexible architecture allows for its use in other areas such as the
simulation of IT systems, queuing networks, hardware architectures or business
processes as well. The Ethernet model used in this benchmark is publicly available
from the address given in the References.” [25]
473.astra
“473.astar (pronounced: A-star) is derived from a portable 2D path-finding
library that is used in game's AI. This library implements three different pathfinding algorithms: First is the well-known A* algorithm for maps with passable and
non-passable terrain types. Second is a modification of the A* path finding algorithm
for maps with different terrain types and different move speed. Third is an
implementation of A* algorithm for graphs. This is formed by map regions with
neighborhood relationship. The library also includes pseudo-intellectual functions
for map region determination.” [25]
483.xalancbmkg
“This program is a modified version of Xalan-C++, an XSLT processor written
in a portable subset of C++. You use the XSLT language to compose XSL style sheets.
An XSL style sheet contains instructions for transforming XML documents from one
document type to another document type (XML, HTML, or other). In structural
terms, an XSL style sheet specifies the transformation of one tree of nodes (the XML
input) into another tree of nodes (the output or transformation result).” [25]
33
3.2.2. CFP2006
410.bwaves
“410.bwaves numerically simulates blast waves in three dimensional
transonic transient laminar viscous flow.
The initial configuration of the blast waves problem consists of a high
pressure and density region at the center of a cubic cell of a periodic lattice, with
low pressure and density elsewhere. Periodic boundary conditions are applied to
the array of cubic cells forming an infinite network. Initially, the high pressure
volume begins to expand in the radial direction as classical shock waves. At the
same time, the expansion waves move to fill the void at the center of the cubic cell.
When the expanding flow reaches the boundaries, it collides with its periodic
images from other cells, thus creating a complex structure of interfering nonlinear
waves. These processes create a nonlinear damped periodic system with energy
being dissipated in time. Finally, the system will come to an equilibrium and steady
state.
The algorithm implemented is an unfactored solver for the implicit solution
of the compressible Navier-Stokes equations using the Bi-CGstab algorithm, which
solves systems of non-symmetric linear equations iteratively.” [26]
416.gamess
34
“A wide range of quantum chemical computations are possible using
GAMESS. The benchmark 416.gamess does the following computations for the
reference workload:



Self-consistent field (SCF) computation (type: Restricted Hartree-Fock) of
cytosine molecule using the direct SCF method
SCF computation (type: Restricted open-shell Hartee-Fock) of water and
cu2+ using the direct SCF method
SCF computation (type: Multi-configuration Self-consisted field) of triazolium
ion using the direct SCF method” [26]
433.milc
“The program generates a gauge field, and is used in lattice gauge theory
applications involving dynamical quarks. Lattice gauge theory involves the study of
some of the fundamental constituents of matter, namely quarks and gluons. In this
area of quantum field theory, traditional perturbative expansions are not useful.
Introducing a discrete lattice of space-time points is the method of choice.” [26]
434.zeusmp
“434.zeusmp is based on ZEUS-MP, a computational fluid dynamics code
developed at the Laboratory for Computational Astrophysics (NCSA, University of
Illinois at Urbana-Champaign) for the simulation of astrophysical phenomena. ZEUSMP solves problems in three spatial dimensions with a wide variety of boundary
conditions.
The program solves the equations of ideal (non-resistive), non-relativistic,
hydrodynamics
and
magnetohydrodynamics,
including
externally
applied
gravitational fields and self-gravity. The gas can be adiabatic or isothermal, and the
35
thermal pressure is isotropic. Boundary conditions may be specified as reflecting,
periodic, inflow, or outflow.” [26]
435.gromacs
“435.gromacs is derived from GROMACS, a versatile package that performs
molecular dynamics, i.e. simulation of the Newtonian equations of motion for
systems with hundreds to millions of particles.
The benchmark version performs a simulation of the protein Lysozyme in a
solution of water and ions. The structure of a protein is normally determined by
experimental techniques such as X-ray crystallography of NMR spectroscopy. By
simulating the atomic motions of these structures, one can gain significant
understanding of protein dynamics and function, and, in some cases, it might even
be possible to predict the structure of new proteins.” [26]
436.cactusADM
“CactusADM is a combination of Cactus, an open source problem solving
environment, and BenchADM, a computational kernel representative of many
applications in numerical relativity (ADM stands for ADM formalism developed by
Arnowitt, Deser and Misner). CactusADM solves the Einstein evolution equations,
which describe how space-time curves as response to its matter content, and are a
set of ten coupled nonlinear partial differential equations, in their standard ADM
3+1 formulation. A staggered-leapfrog numerical method is used to carry out the
update.” [26]
36
437.leslie3d
“437.leslie3d is derived from LESlie3d (Large-Eddy Simulations with LinearEddy Model in 3D), a research-level Computational Fluid Dynamics (CFD) code. It is
the primary solver used to investigate a wide array of turbulence phenomena such
as mixing, combustion, acoustics and general fluid mechanics.
For CPU2006, the program has been set up to solve a test problem which
represents a subset of such flows, namely the temporal mixing layer. This type of
flow occurs in the mixing regions of all combustors that employ fuel injection
(which is nearly all combustors). Also, this sort of mixing layer is a benchmark
problem used to understand physics of turbulent mixing.” [26]
444.namd
“The 444.namd benchmark is derived from the data layout and inner loop of
NAMD, a parallel program for the simulation of large biomolecular systems.
Although NAMD was a winner of a 2002 Gordon Bell award for parallel scalability,
serial performance is equally important to the over 10,000 users who have
downloaded the program over the past several years. Almost all of the runtime is
spent calculating inter-atomic interactions in a small set of functions. This set was
separated from the bulk of the code to form a compact benchmark for CPU2006.
This computational core achieves good performance on a wide range of machines,
but contains no platform-specific optimizations.” [26]
37
447.dealll
“The SPEC CPU2006 benchmark 447.dealII is a program that uses deal.II, a
C++ program library targeted at adaptive finite elements and error estimation. The
library uses state-of-the-art programming techniques of the C++ programming
language, including the Boost library. It offers a modern interface to the complex
data structures and algorithms required for adaptivity and enables use of a variety
of finite elements in one, two, and three space dimensions, as well as timedependent problems.
The main aim of deal.II is to enable development of modern finite element
algorithms, using among other aspects sophisticated error estimators and adaptive
meshes. Writing such programs is a non-trivial task, and successful programs tend
to become very large and complex.” [26]
450.soplex
“450.soplex is based on SoPlex Version 1.2.1. SoPlex solves a linear program
using the Simplex algorithm.” [26]
453.povray
“POV-Ray is a ray-tracer. Ray-tracing is a rendering technique that calculates
an image of a scene by simulating the way rays of light travel in the real world but it
does so backwards. In the real world, rays of light are emitted from a light source
and illuminate objects. The light reflects off of the objects or passes through
transparent objects. This reflected light hits the human eye or a camera lens. As the
38
vast majority of rays never hit an observer, it would take forever to trace a scene.
Thus, ray-tracers like POV-Ray start with their simulated camera and trace rays
backwards out into the scene. The user specifies the location of the camera, light
sources, and objects as well as the surface textures and their interiors.” [26]
454.calculix
“454.calculix is based on CalculiX, which is a free software finite element
code for linear and nonlinear three-dimensional structural applications. It uses the
classical theory of finite elements described in books such as the work by O.C.
Zienkiewicz and R.L. Taylor, "The Finite Element Method", Fourth Edition, McGraw
Hill, 1989. CalculiX can be used to solve a variety of problems such as static
problems (bridge and building design), buckling, dynamic applications (crash,
earthquake resistance) and eigenmode analysis (resonance phenomena).” [26]
459.GemsFDTD
“GemsFDTD solves the Maxwell equations in 3D in the time domain using the
finite-difference time-domain (FDTD) method. The radar cross section (RCS) of a
perfectly conducting (PEC) object is computed. GemsFDTD is a subset of the code
GemsTD developed in the General ElectroMagnetic Solvers (GEMS) project.” [26]
465.tonto
“Tonto is an open source quantum chemistry package, designed by Dylan
Jayatilaka and Daniel J. Grimwood. Objectives include simplicity and portability;
39
aspects not seen in many quantum chemistry codes. The code is easily extendable
by chemists with limited programming skills and time, and is easy to understand
and use.
Tonto is written within an object oriented design, in Fortran 95. It uses
derived types and modules to represent classes. Classes range from integers and
text files, through to atoms, space groups and molecules. There is a "self" variable in
most routines, which should be familiar from many OO languages. Tonto uses
dynamic memory instead of common blocks, and uses array operations where
possible.” [26]
470.lbm
“This program implements the so-called "Lattice Boltzmann Method" (LBM)
to simulate incompressible fluids in 3D as described in. It is the computationally
most important part of a larger code which is used in the field of material science to
simulate the behavior of fluids with free surfaces, in particular the formation and
movement of gas bubbles in metal foams. For benchmarking purposes and easy
optimization for different architectures, the code makes extensive use of macros
which hide the details of the data access. A visualization of the results of the
submitted code can be seen below (flow through a porous medium, grid size
150x150x150, 1000 time steps).” [26]
481.wrf
40
“481.wrf is based on the Weather Research and Forecasting (WRF) Model,
which is a next-generation mesocale numerical weather prediction system designed
to serve both operational forecasting and atmospheric research needs.
WRF features multiple dynamical cores, a 3-dimensional variational (3DVAR)
data assimilation system, and a software architecture allowing for computational
parallelism and system extensibility. The parallel portions of the code have been
turned off for SPEC CPU2006 as the interest here is in single processor
performance.” [26]
482.sphinx3
“Sphinx-3 is a widely known speech recognition system from Carnegie
Mellon University. The 482.sphinx3 benchmark focuses on the CPU-intensive
portion of this speech recognition system.” [26]
3.3.
Rodinia Suite
The following benchmarks were taken from the Rodinia Suite.
Leukocyte
The leukocyte application detects and tracks rolling leukocytes (white blood
cells) in in vivo video microscopy of blood vessels. The velocity of rolling leukocytes
provides important information about the inflammation process, which aids
biomedical researchers in the development of anti-inflammatory medications.
41
In the application, cells are detected in the first video frame and then tracked
through subsequent frames. Detection is accomplished by computing for every pixel
in the frame the maximal Gradient Inverse Coefficient of Variation (GICOV) score
across a range of possible ellipses. The GICOV score for an ellipse is the mean
gradient magnitude along the ellipse divided by the standard deviation of the
gradient magnitude. The matrix of GICOV scores is then dilated to simplify the
process of finding local maxima. For each local maximum, an active contour
algorithm is used to more accurately determine the shape of the cell. [27]
LU Decomposition
LU Decomposition is an algorithm to calculate the solutions of a set of linear
equations. The LUD kernel decomposes a matrix as the product of a lower triangular
matrix and an upper triangular matrix. [27]
SRAD
SRAD (Speckle Reducing Anisotropic Diffusion) is a diffusion method for
ultrasonic and radar imaging applications based on partial differential equations
(PDEs). It is used to remove locally correlated noise, known as speckles, without
destroying important image features. SRAD consists of several pieces of work: image
extraction, continuous iterations over the image (preparation, reduction, statistics,
computation 1 and computation 2) and image compression. The sequential
dependency between all of these stages requires synchronization after each stage
(because each stage operates on the entire image). [27]
42
K-means
K-means are a statistical analysis of clusters where each value is placed into a
group where it has the most similar mean. Initial data points are chosen and the
subsequent iterations shuffle the data around until there is convergence. K-means
is often considered a complex calculation, but over time modified algorithms have
improved speed. [27]
Heart Wall
The heart wall benchmark uses several types of processing to create a
benchmark testing “braided parallelism”. By using image despeckling and edge
detection a new image is produced to detect shapes. Once this process is complete
ellipses are added to the process and finally the entire image is tracked from frame
to frame.
This allows for testing parallelism of both multiple tasks and massive
amounts of data. [27]
Hot Spot
This benchmark runs a simulation of processor power and temperature and
how cells affect their neighbors.
These calculations are done by a series of
differential equations. The differential equations are run on a temperature map by
taking power usage into account until the entire map has been normalized. [27]
Needleman
43
This benchmark is a nonlinear optimization of DNA. It finds the optimal path
to particular cells. Based on cells surrounding elements, initially filled by the
program, backwards calculation is done to find proper alignment. The larger the
calculated score the closer to a match there is. Scores are calculated by looking at
the north, west, and north western cells and points deducted to missing elements.
[27]
Particle Filter
This benchmark estimated locations in noisy environments by looking at the
location and path of an object. This is done by making guesses, checking their
probability, normalizing the guesses, and updating the location of the object. This
implementation of a Particle Filter looking at the speed up provided by GPU
parallelism in order to make this application possible in real time applications. For
our purposes, trials were run on 1, 2 and 5 million data points in 16, 32 and 64 sized
processing grids. [27]
3.4.
John Burkardt Benchmarks
John Burkardt is a computer programmer that has been working in the
computer science field for many years.
His first job was at the Pittsburgh
Supercomputing Company from 1988 – 1992. Up until current day, he has had a
robust career and worked at several places including Bell Helicopter, the
Mathematics Department at Iowa State University, Virginia Tech, and is currently
working at the Department of Scientific Computing at Florida State University. Mr.
44
Burkardt has written many applications, mostly for educational purposes, and two
of them were used in this project for their parallel computing attributes using the
OpenMP library.
These two applications are an FFT benchmark and a Prime
Number Counting benchmark. [28]
FFT
This benchmark is a C program that computes a Fast Fourier Transform
using parallelism via OpenMP. Included in this program is the ability to change the
number of threads used. This allows the user to compare the execution time of
different numbers of threads. [28]
Prime Number Counting
This benchmark is a C program that counts the number of primes between 1
and N. In the default case, N is set to 131,072 but can be changed in the source code.
For the sake of this project, we only used the default case. This program also
allowed the user to change the number of active threads running for comparison of
execution times. [28]
3.5.
SHOC Suite
Scalable
Heterogeneous
Computing
Benchmark
(SHOC)
provides
benchmarks that are setup to run across GPUs, CPUs, and cluster computing through
Message Passing Interface (MPI). While originally designed for use with OpenCL,
the current version 1.1.2 supports NVIDIA’s CUDA language and MPI.
SHOC
45
provides both stress and performance tests, coving areas from mathematical
problem and linear algebra to image processing.
Lennord Jones
This
benchmark,
based
on
molecular
dynamics,
performs
nbody
computations using the Lennord Jones potential. N atoms are spread randomly over
a cubic domain for this particular execution. [29]
Reduction
As the name suggests, this benchmark is massive sum reduction performed
on floating point data. Sets of data points are reduced in individual threads before
being looked over again until a single number is reached. [29]
Chemical Modeling (S3D)
Using a three dimensional grid, with one thread per point, this benchmark is
a massively parallel calculation of chemical rates using S3D.
numerical solver based on the Navier-Stokes equation.
S3D is a direct
This benchmark relies
heavily on floating point calculations, an estimated 10 kFLOPS per thread. [29]
Parallel Prefix Sum
Also known as a scan, this benchmark uses the addition of previous smaller
summations to yield a new overall result. With the addition of multiple parallel
threads, increasing the number of starting values make this harder to run on
modern CPU based systems. [29]
46
Radix Sort
This benchmark is a radix sorting algorithm where sorting is performed
based on individual locations within a number. For example sorting can be based
first on the highest digit, such as the hundreds place for example, and then proceed
to look at the next level down. Sorting as the algorithm processes each level allows
for the list to be sorted quickly in parallel. [29]
2D Stencil Operation
A 9 point stencil computation where each value is updated in turn based on
the X-point pattern defined by the algorithm. The standard pattern that updates
this algorithm remains the same during this benchmark; however, multiple
iterations show speedup in larger systems where multiple points can be processed
at once. [29]
Vector Dot Product
This benchmark calculates and checks the bandwidth of a given device while
performing a vector dot product operation.
Vector dot product results in
calculations to determine of the vectors are orthogonal by performing normalization
and then looking at the angle between the two vectors. [29]
Device Information
47
These benchmarks include calculating data about the particular card that is
being used to execute other benchmarks. These statistics include the download
speed of the device, the devices memory capabilities, and the max flops possible on
the device. Since these numbers could all be limiting factors in the calculation of the
performance of any one device it is important to know where these numbers lie so
that they can be taken into account. [29]
3.6.
Parboil Suite
Parboil is a suite designed by the IMPACT group, Illinois Microarchitecture
Project utilizing Advanced Compiler Technology, mainly for GPUs. Designed to focus
on the massively parallelism systems of today’s GPUs, modeling and complex
mathematics form the basis of the benchmarks. While the IMPACT group also
provides a CPU simulation compiler for CUDA, it is used solely for GPU testing in the
context of our project.
Breadth First Search
Breadth First Searching a style of graph searching algorithm that goes
through every node in a tree until the correct result is found. This algorithm
basically looks through the tree layer by layer, finding children for its next round as
it goes. It will not skip to any of the found children until every node on the current
level is finished being examined. With this type of exhaustive searching large trees
can require large amounts of space and time to compute. Benchmarking with a
48
breadth first searching algorithm will allow us to see the advantages, if any, to more
modern multiprocessor multithreaded systems. [30]
Dense Matrix Multiplication
In matrix multiplication, the dot product of a column by a row is performed.
The limitations on this idea are that the two matrices must have similar dimensions,
or at least have the x dimension of A equal to the y dimension of B. Dense matrix
multiplication follows this same principle, except that the properties of a dense
matrix are different than those of a normal matrix. Consider for example that every
block in a matrix is connected to every other block in the matrix, so that if you
change one you affect the entire matrix. This make multiplication of dense matrices
a time consuming task for simple non parallel processors and an excellent way to
benchmark the idea of parallel processing in modern hardware. [30]
FFT
The Fast Fourier Transform (FFT) is a very important algorithm for Digital
Signal Processing (DSP). Simply put, FFT is a computationally efficient method for
calculating the Discrete Fourier Transform (DFT).
The DFT is a mathematical
operation of transforming a time domain function of finite (discrete) size into a
frequency domain function. There are several ways of computing the DFT, FFT
being one of them. Without FFT, calculating the DFT can be a long and tedious
process. The FFT simplifies this process by breaking the set of data into smaller and
smaller chunks and then calculating the DFT. For example, say there were 32 data
49
points. Trying to calculate the DFT on all 32 points at once would be very difficult.
FFT breaks down that data set into smaller and smaller sets until, in this case, there
are 16 sets of 2. It is much faster and easier to calculate the DFT at a data size of two
than at a data size of 30, especially when parallel computing is involved and several
of those 16 calculations can be done at the same time.
Sum of Absolute Differences
Sum of absolute differences (SAD) is an algorithm often used in video
encoding as it is a comparison between the current image and the next occurring
frame. The algorithm simple finds the absolute value of the given image minus the
next before proceeding to add them all together. The smallest set is the closest to
the same image. SAD is often used across multiple platforms for compressing video,
simple animation, and object recognition. [30]
Distance Cutoff Coulombic Potential
This benchmark computes the Coulombic Potential of each grid point in a 3D
matrix. This benchmark relies on the calculation of the unequal dipole forces of a
water molecule. Speedup is shown by splitting up the calculations among parallel
threads. [30]
Saturating Histogram
Using a 2 dimensional matrix, this benchmark calculates a saturated
histogram based on the input data. For our particular set of tests, the input data is
that of a silicon wafer with a Gaussian representation of data. [30]
50
Lattice Boltzman
The Lattice Boltzman benchmark is a fluid dynamics simulation within an
enclosed container.
Within this benchmark, particle collision and general
interaction is calculated using a discrete equation. This method for fluid dynamic
simulation is easily ported to GPU benchmarking for its ease in parallelism. [30]
Sparse Matrix Dense Vector Multiplication
The SPMV benchmark is similar to the dense matrix multiplication in that all
points are related to one another, thus when one point is acted upon it effects the
results of other points. In this case the matrix itself is sparsely populated so there
are not as many points to work with. This particular execution uses JDS format so it
allows for padding with zeros and multiple alignments. [30]
Two Angular Correction Function
The TPACF benchmark performs a statistical analysis of a spatial
distribution, often used when measuring astronomical bodies.
It calculates a
histogram of distances between every set of objects within the data set. Completion
of this benchmark places distances, normally on an exponential curve, within reach
of a straight sloped line. Parallelism allows for processing of a multitude of points at
one time. [30]
51
4.
Results
4.1.
Devices
The following is the exact specifications for the devices that these
benchmarks were tested on. This information includes: clock rates, system memory,
and host operating system among other specifications.
4.1.1. GPU
All of the GPU benchmarks were run on the same system. Since transfer time
and driver commands are issued by the CPU we must look at the whole system. The
test system is running Ubuntu 10.04 with the Linux kernel version 2.6.32-38generic. Other system specifications are as follows:

CPU: AMD Athlon 64 x2 5200+ running at 2.611GHz dual core

3GB of system memory, DDR2 800MHz

GPU 1: NVIDIA GeForce GTX 460
o CUDA Capability 2.1
o 336 CUDA capable cores
o 1024MB of global onboard memory
o GPU clocks are as follows:


Graphics 675MHz

Memory 1800MHz

Processor 1350MHz
GPU 2: NVIDIA GeForce 9800 GTX+
o CUDA Capability 1.1
o 128 CUDA capable cores
o 512MB of global onboard memory
o GPU clocks are as follows:
52


Graphics 738MHz

Memory 1100MHz

Processor 1836MHz
System PCIE version 1.1
4.1.2. CPU
All of the CPU benchmarks were also run on the same system know at
Worcester Polytechnic Institute as the AMAX machine.
This system has the
following specifications:

2x – Intel Xeon X5650 processors running at 1.6 GHz. The processor is
capable by factory default to run at 2.67GHz. These processors have 6 cores
and 12 threads a piece, totaling 12 cores and 24 threads using Hyper
Threading.

The total memory of this machine is 24.6 GB.

The operating system running on this machine is Linux version 2.6.18274.7.1.e15.

The compiler is GCC, G++, GFortran versions 4.3.4.

Intel C++ Compiler XE version 12.0.1.116 build 2010116
4.1.3. FPGA
The FPGA benchmarks were performed using simulations on a Virtex-5
family board. All simulations were performed using the ISIM program through the
Xilinx program. Virtex-5 boards come in many different series and are specialized
for the following applications taken from the Xilinx datasheet:

Virtex-5 LX: High-performance general logic applications

Virtex-5 LXT: High-performance logic with advanced serial connectivity
53

Virtex-5 SXT: High-performance signal processing applications with
advanced serial connectivity

Virtex-5 TXT: High-performance systems with double density advanced
serial connectivity

Virtex-5 FXT: High-performance embedded systems with advanced serial
connectivity [31]
4.2.
All Devices
As a comparison between all three platforms we looked to the Fast Fourier
Transform. The FFT benchmark was chosen since it was easily portable between all
three devices. For the CPU the benchmark was acquired from John Burkardt, while
the GPU one came from the Parboil Suite and the FPGA core came from MATLAB
HDL Coder.
While the benchmark came from different sources, the FFT
implementation was similar. A factor for concern could come from the use of the
different style compilers, making different optimizations to the code. The timings
from this benchmark can be seen here in Table 1.
FFT Benchmark 262,155 Points
Bench
mark
FFT
CPU
2
Threads
4
Threads
76.17
ms
45.41
ms
8
12
Threads Threads
31.63
ms
27.85
ms
24
Thread
s
31.36
ms
GPU
FPGA
Max Threads
Virtex-5
8.13 us
(Execution)
2.59 ms
Table 1 - FFT Results
54
Using these timings it is obvious that the GPU processed the FFT the fastest.
The GPU throughput at an FFT this size is 311 times more than that of the FPGA and
3351 times faster than the CPU Multicore with 12 threads. What is missing from
this data is the transfer time for the GPU to send and receive the data. This time is
69 ms for this size data. With this time added in the GPU actually becomes the
slowest of all three platforms. Now, the FPGA is the fastest at 26.67 times faster
than the GPU and 10.76 times faster than the CPU.
What these results can infer is that depending on the hookup of your GPU to
minimize transfer time you could be better off using another platform. As GPU
transfer times become smaller and smaller, it is evident that in terms of speed, your
best choice will be a GPU.
4.3.
CPU & GPU
In this section, we will discuss how the CPU and GPU compared against each
other.
The three benchmarks that we were able to test across both of these
platforms were all from the Rodinia benchmark suite.
The Leukocyte, LU
Decomposition, and Speckle Reducing Anisotropic Diffusion (SRAD) benchmarks
were the three that were executed and the average timings from three runs can be
seen as a comparison in Table 2 below.
55
Benchmark
CPU
Leukocyte
2
Threads
11.70 s
LU
Decomposition
Speckle
Reduction
242.82
ms
638.14
ms
GPU
4
Threads
6.11 s
8
Threads
3.65 s
12
Threads
2.16 s
130.00
76.40 ms 69.23 ms
ms
415.07
306.41
283.93
ms
ms
ms
Table 2 - CPU & GPU Results
24
Threads
1.67 s
Max
Threads
0.157 s
145.20
ms
492.26
ms
7.38 ms
282.93
ms
Looking at the above data in most cases the GPUs beats out CPUs in raw
processing. Raw processing is something to note because, for the GPU iterations,
data is all moved to the GPU’s global memory before testing. This keeps the GPU
from having to ask for data over the slower system bus.
Looking at the times required for the Leukocyte benchmark there is a 10.6x
speedup between the 24 Hyper threads of the AXAM machine and the CUDA based
GTX 460. Again this is looking at raw processing as even in the benchmark papers it
is stated that “Although the overall kernel executed in slightly less than a second, the
memory allocation and copying overheads add more than eleven seconds to the
overall runtime.” [32]
The ultimate success of CUDA results from the many code optimizations and
compacting the code into a single kernel; thus, removing much of the extra overhead
that decreases speed.
Moving onto the LU Decomposition, it is easy to see that the GPU was much
faster than the CPU running this benchmark.
This is most likely due to the
functionality of LU Decomposition, which is purely linear algebra mathematics.
56
Since this benchmark is purely mathematical based, the 336 cores and consequently
the numerous ALUs of the GPU would be must faster in this type of computation
than the 12 cores of the CPU. The only place that the CPU would be able to make up
for its slower computation would be in the data transfer portion of this benchmark.
In this study though, we are only looking at raw execution and not data transfer, so
the GPU is much faster than the CPU in the computation of the LU Decomposition
benchmark.
The Speckle Reduction benchmark uses mostly partial differential equations
in its computations.
As described in the background section, this benchmark
consists of several, sequential parts, making synchronization of these parts a
necessary feature. This necessity for synchronization is what most likely allowed
the CPU execution time to equal that of the GPU. In the flat out computational
portions of this benchmark, the GPU would most likely beat the CPU due to its sheer
number of available cores. Advancements in CPU pipelining, however, allow the
CPU to begin the next instruction while the current one is executing. This basically
means that the CPU has the ability to synchronize and start its next instruction at
the same time. We believe that is the uniqueness of CPU pipelining that allows it to
compete with the GPU in execution time for this benchmark.
4.4.
Individual Results
4.4.1. GPU Specific
For any GPU today the biggest area of a setback is in the area of data transfer.
Until recently, and still the majority, GPUs were only on dedicated cards requiring a
57
link to the CPU and memory over a data bus. The limitation of these busses to
transfer the computed data severely limits the abilities of GPUs in high performance
computing. While bus speeds have come a long way since the original PCI and AGP
graphics interfaces, even the newest version of PCIE limits speeds to 32 GB/sec
assuming full utilization of every channel both upload and download.
PCIE Version
Raw Bitrate
( GT/sec )
Total
Bandwidth
( GB/sec )
1.0a
2.5
8
1.1
2.5
8
2.0
5
16
2.1
5
16
3.0
8
32
4.0 (theoretical) 16
X
Table 3 - GeForce Specification
Just looking at the specifications for the GeForce GTX460 we can see that its
maximum memory bandwidth is clocked at 115.2 GB/sec which is far in excess of
the 32 GB/sec we have with the current transfer standard. Looking at the SHOC
Benchmark Suite we can view download speeds and the speed of the onboard
memory of the tested cards as well. In the following table it is clear that larger file
sizes allow for use of more bandwidth, but the graphics cards simply are not capable
of utilizing the full bandwidth. The maximum data rate we see below is 3.278
GB/sec whereas the host systems capability was 4 GB/sec in both the upload and
download channels. With small file sizes, such as the 1kB file, the graphics card
barely reaches a tenth of its potential. As the file size increases past the 512kB point,
the time required to move the data increases in a linear fashion with a factor of two.
58
Focusing on the transfer capabilities of new generations of GPUs would allow for
much faster processing of large data sets by significantly reducing transfer times.
Size of Data Chunk
GB / second
Time
(kB)
(milliseconds)
1
0.0968
0.01057
2
0.1919
0. 01067
4
0.3399
0. 01204
8
0.6074
0. 01348
16
1.0987
0.01491
32
1.5839
0.02069
64
2.1322
0.03074
128
2.5777
0.05084
256
2.8825
0.09094
512
3.0872
0.1698
1024
3.1711
0.3306
2048
3.1773
0.6602
4096
3.1769
1.3211
8192
3.2272
2.5995
16384
3.2552
5.1541
32768
3.2357
10.369
65536
3.2696
20.525
131072
3.2733
41.004
262144
3.2781
81.887
524288
3.2783
163.768
Table 4 - Data Size vs. Data Rate
In contrast to moving data to and from the device, moving data around
between the internal memory levels is a faster process. Looking below at Table 5
gives us another look at how slow global memory transfers are in comparison to the
internal capabilities of the GPU. The internal memory movement is on the order of a
hundred to three hundred GB/sec.
59
Read / Write
Memory Block Size Speed (GB/sec)
Local
32
190 / 181
Local
64
288 / 328
Local
128
299 / 388
Local
256
289 / 385
Local
512
277 / 368
Global
32
7.4 / 3.7
Global
64
5.8 / 3.5
Global
128
4.8 / 3.4
Global
256
4.3 / 3.4
Global
512
4.1 / 3.3
Table 5 – Global vs. Local Memory Speeds
As we can see in the table of timings below, from the Parboil Suite of
Benchmarks, large portions of time during any given benchmark is
transferring data to the GPU or the result back to main memory. While
transferring data is not necessarily the bulk of the time, a more efficient
method for transferring data is needed to help speed up the overall
computation times. As seen in
60
P Average - 460
Average - 9800
Percent Diff
0.036806
0.006477667
-0.92775365
-0.38956379
arboil
CUTCP
GPU
Copy
FFT
GPU
Copy
LBM - long
GPU
Copy
LBM - short
GPU
Copy
MM - long
GPU
Copy
SAD
GPU
Copy
SPVM - large
GPU
Copy
SPVM medium
GPU
Copy
SPVM - small
GPU
Copy
TPACF
GPU
Copy
0.037495333
0.006528333
0.000812667
0.069062
30.15880433
0.354906667
93.86147667 51.36472182
0.334240333 -2.99882802
1.006141333
0.340840333
3.124029333
0.332132333
0.008865333
0.010738333
0.010750667 9.61120174
0.010793667 0.256981856
0.001495667
0.115957
0.002197667 19.00722022
0.064151 -28.7638528
0.000219333
0.076647333
0.000319333 18.56435644
0.054017333 -17.3191426
0.000117
0.051963667
0.000114
0.047360667
4.73333E-05
0.049963667
4.93333E-05 2.068965517
0.045377 -4.81081875
1.361817
0.081934
1.343129333
0.051611333
51.2784621
-1.29396043
-1.2987013
-4.6343125
-0.69087015
-22.7058976
Table 6 the smallest transfer time is still on the order of milliseconds (6.528
ms); magnitudes larger than a single cycle of execution for today’s modern CPUs and
GPUs (0.5 nanosec ). An interesting note comes from the comparison of the older
9800 card to the current 460; the copy times actually increased by an average of
61
9.18% for the new architecture. This may be attributed to the added memory banks
or new hierarchy.
62
Parboil
CUTCP
GPU
Copy
FFT
GPU
Copy
LBM - long
GPU
Copy
LBM - short
GPU
Copy
MM - long
GPU
Copy
SAD
GPU
Copy
SPVM - large
GPU
Copy
SPVM medium
GPU
Copy
SPVM - small
GPU
Copy
TPACF
GPU
Copy
Average - 460
0.037495333
0.006528333
Average - 9800
Percent Diff
0.036806
0.006477667
-0.92775365
-0.38956379
0.000812667
0.069062
30.15880433
0.354906667
93.86147667 51.36472182
0.334240333 -2.99882802
1.006141333
0.340840333
3.124029333
0.332132333
0.008865333
0.010738333
0.010750667 9.61120174
0.010793667 0.256981856
0.001495667
0.115957
0.002197667 19.00722022
0.064151 -28.7638528
0.000219333
0.076647333
0.000319333 18.56435644
0.054017333 -17.3191426
0.000117
0.051963667
0.000114
0.047360667
4.73333E-05
0.049963667
4.93333E-05 2.068965517
0.045377 -4.81081875
1.361817
0.081934
1.343129333
0.051611333
51.2784621
-1.29396043
-1.2987013
-4.6343125
-0.69087015
-22.7058976
Table 6 – Parboil Suite Timings
Continuing to look at the Parboil Suite, there are other noticeable
improvements with the new Fermi architecture. The full table of results can be seen
in the appendices. Using the new architecture, the time spent on interactions
between the CPU and GPU decreased by a noticeable amount. CPU computation
time decreased by 8.77% and the time the CPU spent handling GPU commands
63
decreased by 2.21%.
These increases show how newer GPUs are capable of
handling more commands by themselves and, while still reliant on CPUs, are moving
towards being able to compute by themselves.
Moving to the data from the Rodinia suite other observations on GPU
processing can be made. Once again, as with the Parboil suite, we can see that the
time taken to move memory around with the newer GTX 460 is still slower. Overall
this effect could be due to the test system’s PCIE bus.
We did not have an
intermediate GPU to test floating point speedup with the new architecture, but using
the 9800 GTX+ we could observe native increase. Looking at the table below we see
that the older GPU beat the new architecture by a very slight margin; an average
increase of -0.97% for the GTX 460. Since the older architecture was optimized for
standard arithmetic and did not support floating point, this result is reasonable.
Rodinia
Average - 460
Particle Float
( non-float )
A
GPU execution
B
GPU execution
C
GPU execution
D
GPU execution
E
GPU execution
F
GPU execution
G
GPU execution
H
GPU execution
Average - 9800
Percent Diff
0.000117667
0.000116889
-0.3315964
0.000129333
0.000129444
0.042936883
0.000126333
0.000123444
-1.15658363
0.000120333
0.000116111
-1.78571429
0.000125333
0.000122111
-1.30220027
0.000121
-0.81967213
0.000123667
0.000119889
-1.55109489
0.000129667
0.000127556
-0.82073434
0.000123
Table 7 - Rodinia Particle Filter Results
64
4.4.2. CPU Results
SPEC CPUINT2006 Results
There were two instances of the CPUINT2006 benchmark suite run on the
CPU for this project. The first run through was done on only a single core and single
thread using the GCC, G++, and GFortran 4.3.4 (GNU) compiler. The second was
done using the Intel C++ Compiler XE. The GNU compiler run through had the auto
parallel feature off, meaning it used only a single thread. The Intel compiler run
through had the auto parallel feature enabled, meaning it was using multicore
parallelism to complete the benchmark. The results of these two runs can as well as
the overall speedup percentages can be seen in the following tables.
Iteration #1 [s] Iteration #2 [s] Iteration #3 [s]
400.perlbench
401.bzip2
403.gcc
429.mcf
445.gobmk
456.hmmer
458.jeng
462.libquantum
464.h264ref
471.omnetpp
473.atar
483.xalancbmk
398.00
582.00
375.00
373.00
510.00
869.00
595.00
508.00
710.00
374.00
494.00
268.00
396.00
582.00
375.00
373.00
511.00
869.00
612.00
506.00
706.00
373.00
495.00
260.00
396.00
581.00
375.00
371.00
511.00
869.00
595.00
506.00
709.00
374.00
494.00
259.00
Table 8 - SPEC CPUINT2006 w/ GCC, G++, GFortran compiler w/o auto parallel
65
Iteration #1 [s] Iteration #2 [s] Iteration #3 [s]
400.perlbench
401.bzip2
403.gcc
429.mcf
445.gobmk
456.hmmer
458.jeng
462.libquantum
464.h264ref
471.omnetpp
473.atar
483.xalancbmk
410.00
539.00
330.00
184.00
600.00
213.00
508.00
12.20
959.00
339.00
350.00
214.00
411.00
539.00
331.00
184.00
601.00
214.00
508.00
14.30
959.00
340.00
349.00
214.00
411.00
538.00
334.00
184.00
607.00
214.00
507.00
13.00
1031.00
339.00
349.00
214.00
Table 9 - SPEC CPUINT2006 run with Intel compiler in auto parallel
Iteration
Speedup
400.perlbench
401.bzip2
403.gcc
429.mcf
445.gobmk
456.hmmer
458.jeng
462.libquantum
464.h264ref
471.omnetpp
473.atar
483.xalancbmk
Average
Increase
Per Benchmark
Total Average
Increase
-3.02%
7.39%
12.00%
50.67%
-17.65%
75.49%
14.62%
97.60%
-35.07%
9.36%
29.15%
20.15%
21.72%
#1 Iteration
Speedup
-3.79%
7.39%
11.73%
50.67%
-17.61%
75.37%
16.99%
97.17%
-35.84%
8.85%
29.49%
17.69%
21.51%
#2 Iteration
Speedup
#2
-3.79%
7.40%
10.93%
50.40%
-18.79%
75.37%
14.79%
97.43%
-45.42%
9.36%
29.35%
17.37%
20.37%
21.20%
Table 10 - Speedup percentages from the GNU run to the Intel run
66
From the tables above, you can see that the multicore, auto parallel run
through generally yielded a quicker execution time than the run through with only a
single thread. There were, however, a few of the benchmarks that actually ran
better using only a single thread. This is probably due to the fact that the work load
of these specific benchmarks was better optimized for only a single thread. Overall,
the multicore run through produced a 21.2% increase in speed over than of single
thread run.
SPEC CFP2006 Results
Once again, there were two instances of the SPEC CPU2006 floating point
suite benchmarks run on the CPU. There were two benchmarks in this suite, bwaves
and wrf, which were left out of these runs because of an invalid run error. Both of
these benchmarks would build successfully, but every time we tried to run them, we
would get this invalid run error that we could not figure out how to fix. Otherwise,
all the other benchmarks in the floating point suite ran fine and there timings can be
seen in the following two tables. The first table shows the timings for the run on the
GCC, G++, and GFortran compiler with the auto parallel feature off (single threaded).
The subsequent table shows the timings of the run on the Intel compiler with the
auto parallel feature on (multi-threaded).
67
Iteration #1 [s] Iteration #2 [s] Iteration #3 [s]
416.gamess
937
940
938
433.milc
479
463
489
435.gromacs
579
579
578
436.cactusADM
1372
1441
1338
437.leslie3d
604
604
603
444.namd
496
497
496
447.dealII
430
429
430
450.soplex
270
270
283
453.povray
236
235
237
454.calculix
1484
1484
1484
459.GemsFDTD
517
517
517
465.tonto
652
649
652
470.lbm
378
379
378
482.sphinx3
632
630
632
434.zeusmp
623
625
625
Table 11 - SPEC CFP2006 w/ GCC, G++, GFortran compiler w/o auto parallel
Iteration #1 [s] Iteration #2 [s] Iteration #3 [s]
416.gamess
1238
1185
1197
433.milc
190
189
190
435.gromacs
485
489
482
436.cactusADM
60.9
53.1
50.3
437.leslie3d
87.7
90.5
95.8
444.namd
457
456
457
447.dealII
293
293
293
450.soplex
296
263
286
453.povray
191
191
190
454.calculix
382
292
375
459.GemsFDTD
119
122
121
465.tonto
469
462
469
470.lbm
49.9
49.9
50.1
482.sphinx3
528
544
514
434.zeusmp
93.9
93
90.8
Table 12 - SPEC CFP2006 run with Intel compiler in auto parallel
Though there were a few benchmarks that produced a negative speed
increase from the single threaded to the multi-threaded run, most benchmarks
68
displayed a significant increase in speed.
The speedup of each benchmarking
iteration, as well as the average total speedup of all the run-throughs, can be seen in
the following table.
Iteration #1
Speedup
-32.12%
60.33%
16.23%
95.56%
85.48%
7.86%
31.86%
-9.63%
19.07%
74.26%
76.98%
28.07%
86.80%
16.46%
84.93%
42.81%
Iteration #2
Speedup
-26.06%
59.18%
15.54%
96.32%
85.02%
8.25%
31.70%
2.59%
18.72%
80.32%
76.40%
28.81%
86.83%
13.65%
85.12%
44.16%
Iteration #3
Speedup
-27.61%
61.15%
16.61%
96.24%
84.11%
7.86%
31.86%
-1.06%
19.83%
74.73%
76.60%
28.07%
86.75%
18.67%
85.47%
43.95%
416.gamess
433.milc
435.gromacs
436.cactusADM
437.leslie3d
444.namd
447.dealII
450.soplex
453.povray
454.calculix
459.GemsFDTD
465.tonto
470.lbm
482.sphinx3
434.zeusmp
Average Increase Per
Benchmark
Total Average Increase
43.64%
Table 13 - Speedup percentages from the GNU run to the Intel run
As you can see, there was an average total speedup of 43.64% for the SPEC
CPU2006 floating point suite. This is over two times the speedup we saw from the
CPU2006 integer suite. We believe that such an increase in speedup is due to the
complexity of floating point operations. A single thread would be able to run
simpler integer operations faster than complex floating point operations. Thus, you
would see a more significant speed increase when using the multi-threaded
69
capabilities of a CPU to calculate floating point operations as opposed to integer
operations.
Rodinia / John Burkardt Results
Each of the Rodinia and Burkardt benchmarks was run five different times at
three iterations a piece using 2, 4, 8, 12, and 24 threads. Since the CPU we were
testing on has two processors, totaling 12 cores, we believed that this range of
thread counts would cover the most practical multithreading situations.
The
following table shows the average execution time for three iteration of each
benchmark on the various thread counts. The first five benchmarks in this table are
from the Rodinia suite and the other two are from Burkardt suite.
2 threads 4 threads 8 threads 12 threads 24 threads
Leukocyte (s)
11.70
6.11
3.65
2.16
1.67
LU Decomposition
(ms)
Speckle Reduction
(ms)
Means (s)
242.82
130.00
76.40
69.23
145.20
638.14
415.07
306.41
283.93
492.26
3.35
3.67
2.91
2.16
1.67
Stream Clusters (s)
47.08
25.18
14.61
11.91
10.40
FFT (ms)
76.17
45.41
31.63
27.85
31.36
Primes (s)
2.04
1.17
0.64
0.44
0.36
Table 14 - Rodinia/Burkardt average execution time on 2, 4, 8, 12, 24 threads
70
As you can see, the execution times change when transitioning to a different
number of threads. The total speedup percentage change between thread counts
can be seen in the following table.
2-4 threads 4-8 threads 8-12 threads 12-24 threads
Leukocyte (s)
LU Decomposition
(ms)
Speckle Reduction
(ms)
kmeans (s)
Stream Clusters (s)
FFT (ms)
Primes (s)
Average Increase
Between Threads
47.78%
46.46%
40.26%
41.23%
40.82%
9.38%
22.69%
-109.74%
34.96%
26.18%
7.34%
-73.37%
-9.55%
46.52%
40.38%
42.65%
35.60%
20.71%
41.98%
30.35%
45.30%
35.14%
25.77%
18.48%
11.95%
31.25%
20.71%
22.69%
12.68%
-12.60%
18.18%
-17.07%
Table 15 - Speedup between thread counts
Several interesting observations can be made from the information in the
table above. First, we saw that three out of the seven benchmarks had a decrease in
speed when transitioning between 12 and 24 threads. This is most likely due to the
hardware limitations of our processor.
At twelve threads, our processor can
dedicate one core to each thread because it has a total of 12 cores. Once we
transition to 24 threads though, 12 of the 24 threads become “virtual” threads that
are implemented at the software level. This basically means that each core is
handling two threads apiece.
While this technology may be good for some
applications, running two threads on a single core can sometimes produce slower
execution times than using a single thread per core.
71
The other interesting observation was that the average speed increase from
2 to 4 and 4 to 8 hovered around a 35% increase while the increase between 8 to 12
dropped down to 20%. This shows that there is a significant speed increase up until
8 threads, but beyond that, the speedup begins to become less prevalent.
4.4.3. FPGA Results
As is evident from the benchmarks run on the FPGA, they do not usually
perform as complex tasks as the CPU and GPU benchmarks show. The applications
they are used for are generally specific and are used to enhance applications for
other processes.
Benchmark Clk
Period
(MHz)
FFT
AES
FIR
FP Mul
FIR Core
101
376
710
550
550
Clk
Cycles
Throughput Delay for
(ns)
valid data
(Clock
Cycles)
1
9.87
12
1
2.66
1
1
1.41
8
9
16.4
9
11
20.0
20
Table 16 - FPGA Results
Delay
(ns)
118
2.66
1.13
46.4
36.4
As the Table above shows, the majority of applications run on the FPGA do
not take much time between outputs, but there is usually a larger delay before the
output is actually available.
All of these benchmarks were designed using a
pipelining implementation. This allowed the FPGA to use the ability to break up
tasks and use internal storage to speed up the overall throughput.
FPGAs are very useful to high performance computing, however on an
application specific basis. The advantage of having a higher processing power
72
compared to CPU and the ability to customize the data transfer method to meet your
needs is great. However, it is vital in today’s computing to find the board that has
the proper computational slices to meet the needs of your application.
73
5.
Future Work
For future projects similar to this one, the main topics to focus on would be a
broader spectrum of benchmarks capable of running across all three platforms as
well as possibly looking into newer technologies, such as an Accelerated Processing
Unit (APU). Benchmarking of clusters or testing a single benchmark using multiple
platforms for speedup would also be of use in future research.
While most of today’s benchmarks lie in the realm of scientific and
mathematical algorithms, it would be beneficial to create cross platform
benchmarks across other types of general processing.
Examples of general
processing benchmarks could be ones that handle word processing, weather
tracking, molecule design, encryption, and data compression, all with the capability
of running on all three platforms.
Newer technologies have allowed designers to put both CPUs and GPUs on
the same die, the concept behind APUs. This decreases data transfer times and
allows for newer instruction sets to incorporate both units. These new devices have
the ability to provide substantial performance increases to the processing world.
With this is mind, it would be beneficial to benchmark these new platforms against
their predecessors.
74
6.
Conclusion
High Performance Computing is a rapidly growing field that will require more
research to understand. The technology surrounding CPUs, GPUs, and FPGAs is still
rapidly evolving and will continue in future years. Benchmarking will be a constant
process to stratify different systems as well as different devices.
With the
information in this report, some light is shed on the processing power for different
applications between CPUs, GPUs, and FPGAs.
Overall, 66 benchmarks were investigated over eight suites and sources to
gather information.
These results are useful to compare the three systems
discussed as well as in comparison with other devices during future studies. The
world of High Performance Computing is a constantly evolving field that will play a
significant role in the years to come in many diverse fields.
75
7.
Appendices
7.1.
Parboil Results 9800 GTX+
Parboil
CUTCP
IO
GPU
Copy
Driver
Compute
CPU
Overlap
LBM - long
IO
GPU
Copy
Driver
Compute
CPU
Overlap
LBM short
IO
GPU
Copy
Driver
Compute
CPU
Overlap
MM - long
IO
GPU
Copy
Driver
Compute
CPU
Overlap
GFLOPS
SAD
IO
GPU
Iteration 1
0.039047
0.036803
0.006495
0.000139
0.199751
0.036803
Iteration
2
Iteration
3
0.047604
0.036811
0.006459
0.000141
0.199162
0.036811
0.042384 0.0430117 0.00431289
0.036804 0.036806 4.3589E-06
0.006479 0.0064777 1.8037E-05
0.000141 0.0001403 1.1547E-06
0.199347
0.19942 0.00030121
0.036804 0.036806 4.3589E-06
0.040944 0.040932
93.829116 93.867864
0.335708 0.333089
77.793763 77.834447
1.016244 1.009388
78.004436 78.050671
Average
SD
0.040948
93.88745
0.333924
77.81375
1.009609
78.00946
0.0409413
93.861477
0.3342403
77.813987
1.011747
78.021522
8.3267E-06
0.02968691
0.00133785
0.02034303
0.00389608
0.02536816
0.00121966
0.00225193
0.00076443
2.7839E-05
0.03010198
0.00075295
0.049419
3.121433
0.331451
0.000734
0.920874
0.124557
0.051679
3.125452
0.332959
0.000679
0.976684
0.125872
0.051344
3.125203
0.331987
0.000699
0.968332
0.12585
0.050814
3.1240293
0.3321323
0.000704
0.9552967
0.1254263
3.084938
0.01076
0.010821
0.000104
0.051372
0.000146
3.099593
0.010734
0.010749
0.000124
0.051629
0.000169
3.09249
0.010758
0.010811
0.000121
0.051597
0.000159
3.0923403 0.00732865
0.0107507 1.4468E-05
0.0107937 3.9004E-05
0.0001163 1.0786E-05
0.0515327 0.00014006
0.000158 1.1533E-05
1.55E-13
1.55E-13
1.55E-13
1.554E-13
1.4673E-16
0.166941
0.00221
0.20198
0.002185
0.198754 0.189225
0.002198 0.0021977
0.0193658
1.2503E-05
76
Copy
0.063778
Driver
0.00003
Compute
0.000727
CPU
0.000042
Overlap
SPVM large
IO
0.142549
GPU
0.000319
Copy
0.053893
Driver
0.000063
Compute
0.004965
CPU
0.000079
Overlap
SPVM - medium
IO
0.024611
GPU
0.000117
Copy
0.047665
Driver
0.000064
Compute
0.002742
CPU
0.00008
Overlap
SPVM - small
IO
0.021669
GPU
0.000051
Copy
0.045435
Driver
0.000043
Compute
0.002261
CPU
0.000054
Overlap
TPACF
IO
1.129532
GPU
1.343114
Copy
0.051403
Driver
0.000093
Compute
0.019158
CPU
0.000137
Overlap
0.064767
0.000031
0.00849
0.000042
0.063908 0.064151 0.00053742
0.000029
0.00003
0.000001
0.007983 0.0057333 0.00434302
0.000041 4.167E-05 5.7735E-07
0.087202
0.000319
0.054192
0.000062
0.004981
0.000079
0.113094
0.00032
0.053967
0.000062
0.004978
0.00008
0.02542
0.000112
0.047027
0.00006
0.002828
0.000075
0.02499 0.025007 0.00040477
0.000113 0.000114 2.6458E-06
0.04739 0.0473607 0.00032001
0.000061 6.167E-05 2.0817E-06
0.00279 0.0027867 4.3097E-05
0.000079 0.000078 2.6458E-06
0.000503
0.000048
0.045306
0.000041
0.002294
0.000052
0.02007 0.0140807 0.01178575
0.000049 4.933E-05 1.5275E-06
0.04539 0.045377 6.5475E-05
0.000041 4.167E-05 1.1547E-06
0.002278 0.0022777 1.6503E-05
0.000052 5.267E-05 1.1547E-06
1.067352
1.343151
0.051871
0.000089
0.0188
0.000129
1.07802
1.343123
0.05156
0.00009
0.019022
0.000131
0.1142817 0.02769261
0.0003193 5.7735E-07
0.0540173 0.00015573
6.233E-05 5.7735E-07
0.0049747 8.5049E-06
7.933E-05 5.7735E-07
1.0916347 0.03325068
1.3431293 1.9296E-05
0.0516113 0.00023819
9.067E-05 2.0817E-06
0.0189933 0.00018071
0.0001323 4.1633E-06
77
7.2.
Parboil Results GTX 460
Parboil
CUTCP
IO
GPU
Copy
Driver
Compute
CPU Overlap
FFT
IO
GPU
Copy
Driver
Compute
CPU Overlap
Histogram
IO
GPU
Copy
Driver
Compute
CPU Overlap
LBM - long
IO
GPU
Copy
Driver
Compute
CPU Overlap
LBM - short
IO
GPU
Copy
Driver
Compute
CPU Overlap
MM - long
IO
GPU
Iteration
1
Iteration
2
Iteration
3
Average
SD
0.024785
0.037782
0.006324
0.000138
0.198731
0.037782
0.026053
0.03735
0.006336
0.000148
0.199476
0.03735
0.026624
0.037354
0.006925
0.000155
0.199039
0.037354
0.025820667
0.037495333
0.006528333
0.000147
0.199082
0.037495333
0.000941257
0.000248269
0.000343576
8.544E-06
0.000374357
0.000248269
0.052376
0.000813
0.074595
0.000047
0.000353
0.000062
0.040751
0.000812
0.066348
0.000048
0.000467
0.000064
0.041094
0.000813
0.066243
0.000047
0.000369
0.000064
0.044740333
0.000812667
0.069062
4.73333E-05
0.000396333
6.33333E-05
0.006614905
5.7735E-07
0.004792006
5.7735E-07
6.17198E-05
1.1547E-06
0.152658 0.156614 0.158582
0.140838 0.140679 0.140491
0.155951333
0.140669333
0.003017083
0.000173702
0.139518 0.139357 0.139178
0.00043 0.000496 0.000457
0.140838 0.140679 0.140491
0.139351
0.000461
0.140669333
0.000170079
3.31813E-05
0.000173702
0.047836
30.16735
0.363583
24.9618
0.961555
25.16581
0.049771
30.14701
0.352151
24.94686
0.996269
25.14708
0.049561
30.16205
0.348986
24.94136
1.013636
25.14502
0.049056
30.15880433
0.354906667
24.95000833
0.990486667
25.152638
0.001061756
0.010548376
0.007678761
0.01057388
0.02651762
0.011457143
0.054089
1.005346
0.345664
0.000834
1.020029
0.125006
0.060643
1.004551
0.336826
0.000976
0.9414
0.123213
0.075741
1.008527
0.340031
0.000836
0.927952
0.112586
0.063491
1.006141333
0.340840333
0.000882
0.963127
0.120268333
0.011103405
0.002103939
0.00447424
8.14125E-05
0.049735203
0.006713225
3.195248 3.154655 3.190209
0.008869 0.008867 0.00886
3.180037333
0.008865333
0.022125664
4.72582E-06
78
Copy
Driver
Compute
CPU Overlap
GFLOPS
SAD
IO
GPU
Copy
Driver
Compute
CPU Overlap
SPVM - large
IO
GPU
Copy
Driver
Compute
CPU Overlap
SPVM - medium
IO
GPU
Copy
Driver
Compute
CPU Overlap
SPVM - small
IO
GPU
Copy
Driver
Compute
CPU Overlap
TPACF
IO
GPU
Copy
Driver
Compute
CPU Overlap
0.010757
0.000109
0.092311
0.000153
1.55E-13
0.010806
0.000132
0.081318
0.000179
1.55E-13
0.010652
0.000116
0.080229
0.000158
1.55E-13
0.010738333
0.000119
0.084619333
0.000163333
1.55371E-13
7.86787E-05
1.17898E-05
0.006683396
1.37961E-05
1.15326E-17
0.213267
0.001498
0.126214
0.000038
0.000981
0.000053
0.153
0.001493
0.102197
0.000035
0.000795
0.00005
0.206935
0.001496
0.11946
0.000035
0.000721
0.000049
0.191067333
0.001495667
0.115957
0.000036
0.000832333
5.06667E-05
0.033118952
2.51661E-06
0.012385771
1.73205E-06
0.00013396
2.08167E-06
0.146448
0.000218
0.084568
0.000066
0.004543
0.000081
0.148342
0.00022
0.075166
0.000068
0.004651
0.000084
0.148184
0.00022
0.070208
0.000066
0.004879
0.000082
0.147658
0.000219333
0.076647333
6.66667E-05
0.004691
8.23333E-05
0.001050864
1.1547E-06
0.007293707
1.1547E-06
0.000171534
1.52753E-06
0.023466
0.000118
0.051357
0.000063
0.002235
0.000078
0.024767
0.000116
0.05229
0.000061
0.002791
0.000076
0.024647
0.000117
0.052244
0.000062
0.002609
0.000077
0.024293333
0.000117
0.051963667
0.000062
0.002545
0.000077
0.000719
1E-06
0.000525892
0.000001
0.000283471
1E-06
0.007575
0.000048
0.049176
0.000043
0.001726
0.000055
0.014441
0.000047
0.050157
0.000042
0.002104
0.000054
0.015238
0.000047
0.050558
0.000043
0.00209
0.000054
0.012418
4.73333E-05
0.049963667
4.26667E-05
0.001973333
5.43333E-05
0.00421305
5.7735E-07
0.000710995
5.7735E-07
0.000214311
5.7735E-07
1.18402 1.18196 1.177699
1.361854 1.361822 1.361775
0.095364 0.075245 0.075193
0.000113 0.00011
0.0001
0.00904 0.01928 0.020788
0.000156 0.000152 0.000142
1.181226333
1.361817
0.081934
0.000107667
0.016369333
0.00015
0.003223734
3.97366E-05
0.01163075
6.80686E-06
0.006392015
7.2111E-06
79
7.3.
Rodinia Results 9800 GTX+
Rodinia
Iteration 1
Iteration 2
Iteration 3
LUD
64
256
512
2048
Particle Float (naïve)
A
send from GPU
send to GPU
GPU execution
Total
B
send from GPU
send to GPU
GPU execution
Total
C
send from GPU
send to GPU
GPU execution
Total
D
send from GPU
send to GPU
GPU execution
Total
E
send from GPU
send to GPU
GPU execution
Total
F
send from GPU
send to GPU
GPU execution
Total
G
send from GPU
send to GPU
ms
Average
SD
0.238
1.185
3.311
34.445
0.239
1.512
3.315
34.534
0.239
1.456
3.312
34.511
0.238666667
1.384333333
3.312666667
34.49666667
0.00057735
0.17488377
0.00208167
0.04619885
0.02881
0.037435
0.00016
3.284789
0.034164
0.037895
0.000167
3.280237
0.032453
0.031809 0.00273448
0.03769 0.037673333 0.00023045
0.000166 0.000164333 3.7859E-06
3.28295 3.282658667 0.00228994
0.059425
0.074129
0.000162
6.524812
0.060929
0.074013
0.000164
6.702561
0.060115 0.060156333 0.00075285
0.074069 0.074070333 5.8011E-05
0.000164 0.000163333 1.1547E-06
6.593282
6.606885 0.08965187
0.146823
0.186539
0.000173
16.691667
0.149165
0.184094
0.000169
16.667685
0.029974
0.037564
0.000154
3.278011
0.029582
0.037992
0.000159
3.289114
0.029834 0.029796667 0.00019865
0.0376982
0.0377514 0.0002189
0.000156 0.000156333 2.5166E-06
3.283495
3.28354 0.00555164
0.058106
0.074384
0.000167
6.527779
0.056998
0.073757
0.000159
6.501522
0.057398 0.057500667 0.00056109
0.074287 0.074142667 0.0003375
0.000166
0.000164 4.3589E-06
6.51213 6.513810333 0.0132089
0.141125
0.184281
0.000169
16.157624
0.139959
0.182941
0.000166
16.205709
0.13997 0.140351333 0.00067004
0.183213 0.183478333 0.00070831
0.000168 0.000167667 1.5275E-06
16.199371
16.187568 0.02612518
0.029478
0.037714
0.029737
0.037855
0.029587 0.029600667 0.00013004
0.037729
0.037766
7.744E-05
sec
0.148302
0.185983
0.00017
16.682983
0.148096667 0.00118442
0.185538667 0.00128163
0.000170667 2.0817E-06
16.68077833 0.01214205
80
GPU execution
Total
H
send from GPU
send to GPU
GPU execution
Total
7.4.
0.000163
3.286379
0.00016
3.285156
0.000161 0.000161333 1.5275E-06
3.285983 3.285839333 0.00062403
0.058453
0.074478
0.000163
6.700087
0.057608
0.074515
0.000173
6.518035
0.057983 0.058014667 0.00042339
0.745983 0.298325333 0.38768291
0.001064 0.000466667 0.00051733
6.690865
6.636329 0.10254933
Rodinia Results GTX 460
Rodinia
Leukocyte
Detection
computation
dilation
total
Tracking
computation
evolution
total
TOTAL
LUD
64
256
512
2048
Particle Float
(float)
A
send from GPU
send to GPU
GPU execution
Total
B
send from GPU
send to GPU
GPU execution
Total
D
send from GPU
Iteration
1
sec
Iteration
2
Iteration
3
Average
SD
0.0185
0.01006
0.08415
sec
0.04268
0.01027
0.0655
4.01395
ms
0.575
2.848
7.374
115.623
sec
0.01944
0.01068
0.09912
0.01946
0.01068
0.09981
0.019133333
0.010473333
0.09436
0.000548574
0.000357957
0.008848847
0.04274
0.01027
0.06538
4.02195
0.04282
0.01037
0.06544
4.02639
0.042746667
0.010303333
0.06544
4.020763333
7.02377E-05
5.7735E-05
6E-05
0.006304327
0.579
2.828
7.388
115.375
0.575
2.83
7.387
117.59
0.576333333
2.835333333
7.383
116.196
0.002309401
0.011015141
0.00781025
1.213590953
0.484112 0.48184 0.480469
0.014158 0.013941 0.013862
0.000251 0.000258 0.000255
0.65792 0.584687 0.61418
0.482140207
0.013987
0.000254667
0.618929
0.001840149
0.000153268
3.51188E-06
0.036846748
0.977968 0.979106 0.973441
0.028367 0.027495 0.027456
0.000267 0.00024 0.000237
1.099265 1.099386 1.094044
0.976838333
0.027772667
0.000248
1.097565
0.002996693
0.000515077
1.65227E-05
0.003049876
21.25275 14.00893 30.37887
21.88018267
8.202989064
81
send to GPU
GPU execution
Total
E
send from GPU
send to GPU
GPU execution
Total
G
send from GPU
send to GPU
GPU execution
Total
H
send from GPU
send to GPU
GPU execution
Total
Particle Float
(naïve)
A
send from GPU
send to GPU
GPU execution
Total
B
send from GPU
send to GPU
GPU execution
Total
C
send from GPU
send to GPU
GPU execution
Total
D
send from GPU
send to GPU
GPU execution
Total
E
send from GPU
send to GPU
0.013946 0.013852 0.013902
0.000265 0.00253 0.000255
21.34469 14.10073 30.46716
0.0139
0.001016667
21.97086
4.70319E-05
0.001310595
8.201161321
8.556762
0.027564
0.000234
8.675124
8.546939
0.027571
0.000244
8.666531
8.539814
0.027614
0.000259
8.660525
8.547838333
0.027583
0.000245667
8.667393333
0.008509717
2.7074E-05
1.25831E-05
0.007337603
32.85077
0.013938
0.000251
32.94073
13.84569 14.8465
0.013869 0.013931
0.000245 0.000248
13.93179 14.93657
20.51432133
0.013912667
0.000248
20.603027
10.69539136
3.79781E-05
3E-06
10.69656602
8.535585
8.5487 8.544959
0.027881 0.027692 0.027521
0.000245 0.000236 0.000241
8.656592 8.669315 8.670502
sec
8.543081333
0.027698
0.000240667
8.665469667
0.006756111
0.000180075
4.50925E-06
0.007711159
18.03987 11.67335 11.65105
0.037767 0.037449 0.037533
0.00012 0.00012 0.000113
21.40222 14.9382 14.92392
13.78808867
0.037583
0.000117667
17.08811367
3.682170995
0.000164791
4.04145E-06
3.736133375
8.695787
0.075593
0.000129
15.20682
9.222604 9.195602
0.075548
0.0736
0.000118 0.000141
15.67589 15.67246
9.037997667
0.074913667
0.000129333
15.518391
0.296670494
0.001137891
1.15036E-05
0.269830416
14.00524
0.187186
0.000135
30.2434
8.159942
0.183569
0.000118
24.24978
8.216452
0.185398
0.000126
24.24987
10.12720967
0.185384333
0.000126333
26.247683
3.358587309
0.001808539
8.5049E-06
3.460390697
12.49726
0.038156
0.000133
15.77098
12.42138
0.037788
0.000117
15.77281
12.48212
0.037667
0.000111
15.74104
12.46691867
0.037870333
0.000120333
15.76160833
0.040157217
0.000254685
1.13725E-05
0.017839721
10.85223 10.2257 9.016069
0.075725 0.075797 0.07437
10.031332
0.075297333
0.933383694
0.000803901
82
GPU execution
Total
F
send from GPU
send to GPU
GPU execution
Total
G
send from GPU
send to GPU
GPU execution
Total
H
send from GPU
send to GPU
GPU execution
Total
0.000135 0.000117 0.000124
46.67132 16.67096 15.6721
0.000125333
26.33812367
9.07377E-06
17.6161456
14.11983 15.16998 14.26351
0.18705 0.185197 0.183663
0.000129 0.00012 0.00012
30.24967 31.24654 30.24503
14.51777167
0.185303333
0.000123
30.58041333
0.56937989
0.001696002
5.19615E-06
0.576885553
13.49624
0.037861
0.000135
16.7578
12.51785
0.037663
0.000111
15.77775
12.84216133
0.037726667
0.000123667
16.102446
0.566457901
0.000116389
1.20554E-05
0.567560183
9.182923 10.20361 9.200904
0.075046 0.075518 0.074179
0.000136 0.000123 0.00013
15.67639 16.6652 15.67365
9.529146333
0.074914333
0.000129667
16.00507867
0.584173588
0.000679141
6.50641E-06
0.571680899
12.51239
0.037656
0.000125
15.77179
83
7.5.
SHOC Max Flops GTX 460
84
7.6.
SHOC Bus Download Speed GTX 460
85
7.7.
SHOC Device Memory GTX 460
86
7.8.
SHOC SPMV GTX 460
87
7.9.
SHOC MD GTX 460
88
7.10. SHOC Reduction GTX 460
89
7.11. SHOC S3D GTX 460
90
7.12. SHOC Scan GTX 460
91
7.13. SHOC SGEMM GTX 460
92
7.14. SHOC Sort GTX 460
93
7.15. SHOC Stencil 2D GTX 460
94
7.16. SHOC Triad GTX 460
95
7.17. SHOC Max Flops 9800 GTX+
96
7.18. SHOC Bus Download Speed 9800 GTX+
97
7.19. SHOC SPMV 9800 GTX+
98
7.20. SHOC MD 9800 GTX+
99
7.21. SHOC Reduction 9800 GTX+
100
7.22. SHOC S3D 9800 GTX+
101
7.23. SHOC SGEMM 9800 GTX+
102
7.24. SHOC Sort 9800 GTX+
103
7.25. SHOC Stencil 2D 9800 GTX+
104
7.26. SHOC Triad 9800 GTX+
105
7.27. SPEC CPU2006 Integer Results (No Auto-Parallel)
Iteration #1 [s] Iteration #2 [s] Iteration #3 [s]
400.perlbench
401.bzip2
403.gcc
429.mcf
445.gobmk
456.hmmer
458.jeng
462.libquantum
464.h264ref
471.omnetpp
473.atar
483.xalancbmk
398.00
582.00
375.00
373.00
510.00
869.00
595.00
508.00
710.00
374.00
494.00
268.00
396.00
582.00
375.00
373.00
511.00
869.00
612.00
506.00
706.00
373.00
495.00
260.00
396.00
581.00
375.00
371.00
511.00
869.00
595.00
506.00
709.00
374.00
494.00
259.00
7.28. SPEC CPU2006 Integer Results (Auto-Parallel Enabled)
Iteration #1 [s] Iteration #2 [s] Iteration #3 [s]
400.perlbench
401.bzip2
403.gcc
429.mcf
445.gobmk
456.hmmer
458.jeng
462.libquantum
464.h264ref
471.omnetpp
473.atar
483.xalancbmk
410.00
539.00
330.00
184.00
600.00
213.00
508.00
12.20
959.00
339.00
350.00
214.00
411.00
539.00
331.00
184.00
601.00
214.00
508.00
14.30
959.00
340.00
349.00
214.00
411.00
538.00
334.00
184.00
607.00
214.00
507.00
13.00
1031.00
339.00
349.00
214.00
106
7.29. Speedup of SPEC CPU2006 Integer Results
Iteration
Speedup
400.perlbench
401.bzip2
403.gcc
429.mcf
445.gobmk
456.hmmer
458.jeng
462.libquantum
464.h264ref
471.omnetpp
473.atar
483.xalancbmk
Average
Increase
Per Benchmark
Total Average
Increase
-3.02%
7.39%
12.00%
50.67%
-17.65%
75.49%
14.62%
97.60%
-35.07%
9.36%
29.15%
20.15%
21.72%
#1 Iteration
Speedup
#2 Iteration
Speedup
-3.79%
7.39%
11.73%
50.67%
-17.61%
75.37%
16.99%
97.17%
-35.84%
8.85%
29.49%
17.69%
21.51%
#2
-3.79%
7.40%
10.93%
50.40%
-18.79%
75.37%
14.79%
97.43%
-45.42%
9.36%
29.35%
17.37%
20.37%
21.20%
7.30. SPEC CPU2006 Floating Point Results (No Auto-Parallel)
Iteration #1 [s] Iteration #2 [s] Iteration #3 [s]
416.gamess
937
940
938
433.milc
479
463
489
435.gromacs
579
579
578
436.cactusADM
1372
1441
1338
437.leslie3d
604
604
603
444.namd
496
497
496
447.dealII
430
429
430
450.soplex
270
270
283
453.povray
236
235
237
454.calculix
1484
1484
1484
459.GemsFDTD
517
517
517
465.tonto
652
649
652
470.lbm
378
379
378
107
482.sphinx3
434.zeusmp
632
623
630
625
632
625
7.31. SPEC CPU2006 Floating Point Results (Auto-Parallel Enabled)
Iteration #1 [s] Iteration #2 [s] Iteration #3 [s]
416.gamess
1238
1185
1197
433.milc
190
189
190
435.gromacs
485
489
482
436.cactusADM
60.9
53.1
50.3
437.leslie3d
87.7
90.5
95.8
444.namd
457
456
457
447.dealII
293
293
293
450.soplex
296
263
286
453.povray
191
191
190
454.calculix
382
292
375
459.GemsFDTD
119
122
121
465.tonto
469
462
469
470.lbm
49.9
49.9
50.1
482.sphinx3
528
544
514
434.zeusmp
93.9
93
90.8
7.32. Speedup of SPEC CPU2006 Floating Point Results
416.gamess
433.milc
435.gromacs
436.cactusADM
437.leslie3d
444.namd
447.dealII
450.soplex
453.povray
454.calculix
459.GemsFDTD
465.tonto
470.lbm
Iteration #1
Speedup
-32.12%
60.33%
16.23%
95.56%
85.48%
7.86%
31.86%
-9.63%
19.07%
74.26%
76.98%
28.07%
86.80%
Iteration #2
Speedup
-26.06%
59.18%
15.54%
96.32%
85.02%
8.25%
31.70%
2.59%
18.72%
80.32%
76.40%
28.81%
86.83%
Iteration #3
Speedup
-27.61%
61.15%
16.61%
96.24%
84.11%
7.86%
31.86%
-1.06%
19.83%
74.73%
76.60%
28.07%
86.75%
108
482.sphinx3
434.zeusmp
Average Increase Per
Benchmark
Total Average Increase
16.46%
84.93%
42.81%
13.65%
85.12%
44.16%
18.67%
85.47%
43.95%
43.64%
7.33. Rodinia/Burkardt Benchmarks Average Execution Times
2 threads 4 threads 8 threads 12 threads 24 threads
Leukocyte (s)
11.70
6.11
3.65
2.16
1.67
LU Decomposition
(ms)
Speckle Reduction
(ms)
Kmeans (s)
242.82
130.00
76.40
69.23
145.20
638.14
415.07
306.41
283.93
492.26
3.35
3.67
2.91
2.16
1.67
Stream Clusters (s)
47.08
25.18
14.61
11.91
10.40
FFT (ms)
76.17
45.41
31.63
27.85
31.36
Primes (s)
2.04
1.17
0.64
0.44
0.36
7.34. Rodinia/Burkardt Benchmarks Speedup between Thread Count
2-4 threads 4-8 threads 8-12 threads 12-24 threads
Leukocyte (s)
LU Decomposition
(ms)
Speckle Reduction
(ms)
kmeans (s)
Stream Clusters (s)
FFT (ms)
Primes (s)
Average Increase
Between Threads
47.78%
46.46%
40.26%
41.23%
40.82%
9.38%
22.69%
-109.74%
34.96%
26.18%
7.34%
-73.37%
-9.55%
46.52%
40.38%
42.65%
35.60%
20.71%
41.98%
30.35%
45.30%
35.14%
25.77%
18.48%
11.95%
31.25%
20.71%
22.69%
12.68%
-12.60%
18.18%
-17.07%
109
7.35. FPGA Results
Benchmark Clk
Period
(MHz)
Clk
Cycles
FFT
AES
FIR
FP Mul
FIR Core
1
1
1
9
11
101
376
710
550
550
Throughput Delay for
(ns)
valid data
(Clock
Cycles)
9.87
12
2.66
1
1.41
8
16.4
9
20.0
20
Delay
(ns)
118
2.66
1.13
46.4
36.4
110
Bibliography
[1] "The
Methodology
of
Random
Logic
Design,"
[Online].
http://www.intel4004.com/mrld.htm. [Accessed 20 January 2012].
Available:
[2] D.
Risley,
"PCMech,"
23
March
2001.
[Online].
Available:
http://www.pcmech.com/article/a-cpu-history/10/. [Accessed 21 January 2012].
[3] "AMD," 5 June 2000. [Online]. Available: http://www.amd.com/us/pressreleases/Pages/Press_Release_729.aspx. [Accessed 20 January 2012].
[4] C.
S.
a.
D.
Patterson,
"Berkley,"
[Online].
Available:
http://www.eecs.berkeley.edu/Pubs/TechRpts/1982/CSD-82-106.pdf. [Accessed 20
January 2012].
[5] M.
Schmalz,
"University
of
Florida,"
http://www.cise.ufl.edu/~mssz/CompOrg/CDA-proc.html.
2012].
[Online].
[Accessed
Available:
21 January
[6] J. Tyson, "HowStuffWorks.com," 23 August 2000. [Online]. Available:
http://computer.howstuffworks.com/computer-memory4.htm. [Accessed 21 January
2012].
[7] T. Soderstrom, "Tom's Hardware," 11 December 2006. [Online]. Available:
http://www.tomshardware.com/reviews/overclocking-guide-part-1,1379.html.
[Accessed 20 January 2012].
[8] G. Petley, "VlsiTechnology," 22 September 2008.
http://www.vlsitechnology.org/. [Accessed 22 January 2012].
[Online].
Available:
[9] "SeachDataCenter,"
October
2004.
[Online].
Available:
http://searchdatacenter.techtarget.com/definition/multi-core-processor. [Accessed 23
January 2012].
[10] B. Barney, "Lawrence Livermore National Laboratory," [Online]. Available:
https://computing.llnl.gov/tutorials/parallel_comp/. [Accessed 23 January 2012].
[11] NVIDIA,
"What
is
GPU
Computing?,"
2012.
[Online].
Available:
http://www.nvidia.com/object/GPU_Computing.html. [Accessed 18 1 2012].
[12] P. Lilly, "From Voodoo to GeForce: The Awsome History of 3D Graphics," 19 5 2009.
[Online].
Available:
www.maximumpc.com/article/features/voodoo_geforce_awsome_history_3d_graphics.
[Accessed 26 1 2012].
[13] "Graphics
Processing
Unit,"
12
1
2012.
[Online].
Available:
111
en.wikipedia.oeg/wiki/Grpahics_processing_unit. [Accessed 19 1 2012].
[14] NVIDIA, "PTX: Parallel Thread Execution ISA Version 2.3," NVIDIA, 2011.
[15] "NVIDIA CUDA Compute Capability Comparative Table," 6 6 2010. [Online]. Available:
www.geeks3d.com/20100606/gpu-computing-nvidia-cuda-compute-capabilitycomparative-table/. [Accessed 11 12 2011].
[16] NVIDIA, "NVIDIA's Next Generation CUDA Compute Architecture: Fermi v1.1," NVIDIA,
2009.
[17] NVIDIA, "NVIDIA CUDA Architecture," NVIDIA, 2009.
[18] F. E. Allen, "The Greatest Inventors You've Never Heard Of," 2009.
[19] "FPGA
Applications
&
Consulting
Experts,"
[Online].
Available:
http://fpgaace.com/index.php/fpga-history/. [Accessed 27 February 2012].
[20] "FPGA Central," 2011. [Online]. Available: http://www.fpgacentral.com/docs/fpgatutorial/history-programmable-logic. [Accessed 28 January 2012].
[21] V.
Betz,
"University
of
Toronto,"
[Online].
Available:
http://www.eecg.toronto.edu/~vaughn/challenge/fpga_arch.html. [Accessed February
1 2012].
[22] "Electrical
Engineering
Times,"
October
2008.
[Online].
Available:
http://www.eetimes.com/electrical-engineers/educationtraining/courses/4000134/Fundamentals-of-FPGAs. [Accessed 2 February 2012].
[23] B. C. L. a. A. E. Crews, "The Evolution of Benchmarking as a Computer Performance
Evaluation Technique," MIS Quarterly, vol. 9, no. 1, pp. 7-16, 1985.
[24] "SPEC
CPU
2006,"
7
September
2011.
[Online].
http://www.spec.org/cpu2006/. [Accessed 22 January 2012].
Available:
[25] "SPEC,"
24
August
2006.
[Online].
http://www.spec.org/cpu2006/CINT2006/. [Accessed 22 January 2012].
Available:
[26] "SPEC,"
27
September
2006.
[Online].
http://www.spec.org/cpu2006/CFP2006/. [Accessed 22 January 2012].
Available:
[27] "Rodinia: Accelerated Compute-Intensive Applicaitons with Accelerators," 16 12 2011.
[Online].
Available:
https://www.cs.virginia.edu/~skadron/wiki/rodinia/index.php/Main_Page. [Accessed
18 12 2011].
[28] J. Burkardt, "Florida State University," 03 September 2011. [Online]. Available:
http://people.sc.fsu.edu/~jburkardt/. [Accessed 23 January 2012].
[29] "Scalable HeterOgeneous Computing (SHOC) Benchmarking Suite," Oak Ridge National
Laboratory, 11 11 2011. [Online]. Available: ft.ornl.gov/doku/shoc/start. [Accessed 19
112
12 2011].
[30] "Parboil Benchmark Suite," Illinois Microarchitecture Project utilizing Advanced
Compiler
Technology
(IMPACT),
[Online].
Available:
impact.crhc.illinois.edu/parboil.php. [Accessed 18 12 2011].
[31] "Xilinx,"
6
February
2009.
[Online].
Available:
http://www.xilinx.com/support/documentation/data_sheets/ds100.pdf. [Accessed 12
February 2012].
[32] "University
of
Virginia,"
[Online].
Available:
http://www.cs.virginia.edu/!skadron/Papers/boyer_leukocyte_ipdps09.pdf. [Accessed
23 January 2012].
[33] M. Strickland, "HEARST Electronic Products," 1 March 2010. [Online]. Available:
http://www2.electronicproducts.com/The_evolution_of_FPGA_coprocessing-articleFAJH_FPGA_Mar2010-html.aspx. [Accessed January 29 2012].
[34] J. Sanders and E. Kandrot, CUDA by Example, Boston, MA: Pearson Education, 2011.
[35] NVIDIA, "NVIDIA CUDA C Programming Guide," NVIDIA, 2011.
[36] W.-m. Hwu and D. Kirk, Programming Massively Parallel Processors: A Hands-on
Approach, Burlington, MA: Morgan Kaufmann, 2010.
[37] A. Danalis, P. Roth, G. Marin, K. Spafford, C. McCurdy, V. Tipparaju, J. Meredith and J.
Vetter, "The Scalabe HeterOgeneous Computing (SHOC) Benchmark Suite," Pittsburgh,
PA, 2010.
[38] S. Che, M. Boyer, J. Meng, D. Tarjan, J. Sheaffer, S.-H. Lee and K. Skadron, "Rodinia: A
Benchmark Suite for Heterogeneous Computing," Dept Computer Science, UVA, 2009.
[39] F. Abi-Chahla, "A Few Definitions," 18 6 2008. [Online]. Available:
www.tomshardware.com/reviews/nvidia-cuda-gpu,1945-7.html. [Accessed 27 1
2012].
[40] "National
Instruments,"
19
December
2011.
[Online].
Available:
http://zone.ni.com/devzone/cda/tut/p/id/6983. [Accessed 27 January 2012].
113